Merge pull request 'Add web game (plan2.md) on the independent runtime, merge-blocking gates' (#1) from web-game into main

Reviewed-on: #1

.github/rulesets/README.md

@@ -0,0 +1,26 @@
+# Merge enforcement (finding 1 — external dependency)
+
+A workflow file in `.github/workflows/` **defines** jobs; it does **not** enforce
+that they pass before merge. Enforcement is a server-side GitHub setting
+(branch protection / repository ruleset) that marks the jobs as **required
+status checks** on `main` and the merge queue. That setting lives in the GitHub
+repository configuration, not in this repository's tree, and applying it
+requires repository-admin privileges and an authenticated `gh`/API token.
+
+This is therefore BLOCKED on external infrastructure. To close the gap, a repo
+admin applies the ruleset in `main-required-checks.json`:
+
+```bash
+# Requires: gh auth login as a repo admin
+gh api -X POST repos/<owner>/<repo>/rulesets \
+  --input .github/rulesets/main-required-checks.json
+
+# Verify the required checks are active:
+gh api repos/<owner>/<repo>/rulesets --jq '.[].name'
+gh api repos/<owner>/<repo>/branches/main/protection 2>/dev/null \
+  || echo "no classic protection (rulesets in use)"
+```
+
+Until that ruleset is active, the `merge-gates` and `web-rust-gates` jobs are
+*advisory CI*, not merge enforcement. Do not treat their presence in the tree as
+satisfying the merge-blocking requirement.

.github/rulesets/main-required-checks.json

@@ -0,0 +1,27 @@
+{
+  "name": "main-merge-gates",
+  "target": "branch",
+  "enforcement": "active",
+  "conditions": { "ref_name": { "include": ["refs/heads/main"], "exclude": [] } },
+  "rules": [
+    { "type": "pull_request",
+      "parameters": {
+        "required_approving_review_count": 0,
+        "dismiss_stale_reviews_on_push": true,
+        "require_code_owner_review": false,
+        "require_last_push_approval": false,
+        "required_review_thread_resolution": false
+      }
+    },
+    { "type": "required_status_checks",
+      "parameters": {
+        "strict_required_status_checks_policy": true,
+        "required_status_checks": [
+          { "context": "merge-gates" },
+          { "context": "web-rust-gates" }
+        ]
+      }
+    },
+    { "type": "non_fast_forward" }
+  ]
+}

.github/workflows/merge-gates.yml

@@ -0,0 +1,93 @@
+name: magicka-merge-gates
+
+# The merge-blocking enforcement point for the Phase 0/1 acceptance gates.
+# Per the spec's compliance model, acceptance may NOT be satisfied by a locally
+# runnable binary, a default (fast) profile, or a documentation claim. This
+# workflow is the enforcement: the `merge-gates` job below runs the full merge
+# profile (1,000,000 executions, 10 perturbations each, 100% reference/runtime
+# comparison, 10,000 persisted replay cases, >=500 mutants) and must be a
+# REQUIRED status check on the protected branch / merge queue. If it fails or is
+# absent, merge is blocked.
+
+on:
+  pull_request:
+    branches: [main]
+  merge_group:            # GitHub merge queue — the merge-blocking entry point
+  push:
+    branches: [main]
+
+jobs:
+  # Advisory PR feedback only. The spec permits a 10% fast slice for PR signal,
+  # but this job is explicitly NOT acceptance and never substitutes for the
+  # merge profile.
+  advisory-fast:
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+      - name: Unit tests (incl. negative controls)
+        run: cargo test --release --workspace
+      - name: Fast advisory gate run
+        env:
+          MAGICKA_PROFILE: fast
+        run: cargo run --release -p ci_reports --bin ci
+
+  # The acceptance gate. Required on merge_group / protected main.
+  merge-gates:
+    if: github.event_name == 'merge_group' || github.event_name == 'push'
+    runs-on: ubuntu-latest
+    timeout-minutes: 1440
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Verify committed replay corpus has >= 10,000 cases
+        run: |
+          lines=$(grep -vcE '^(#|master)' crates/replay_corpus/corpus/replay_corpus.tsv)
+          echo "replay corpus cases: $lines"
+          test "$lines" -ge 10000
+
+      - name: Unit tests (incl. negative controls)
+        run: cargo test --release --workspace
+
+      # plan2.md Phase H: the protocol + socket + replay/visibility gates are
+      # merge-blocking. These run the full HTTP/WebSocket/protocol/runtime path
+      # headlessly over real sockets (1k matches, 10k fuzz, 100 e2e, leak +
+      # resilience). The rendered-browser layer is NOT gated here — see
+      # web-gates.yml (advisory, blocked on CI infrastructure).
+      - name: Web protocol + socket gates (merge-blocking)
+        run: cargo test --release -p protocol -p game_runtime -p server -p web_assets -p web_client -p web_tests
+
+      - name: Full merge-blocking acceptance gates
+        env:
+          MAGICKA_PROFILE: merge
+          MAGICKA_OUT: ci_out
+        run: cargo run --release -p ci_reports --bin ci
+
+      - name: Enforce required artifacts exist
+        run: |
+          for r in domain_participation_report causal_rank_report \
+                   causal_explanation_report \
+                   compression_resistance_report metamorphic_response_report \
+                   mutation_survivor_report runtime_equivalence_report \
+                   replay_report coverage_report provenance_report \
+                   compliance_report; do
+            test -s "ci_out/${r}.json" || { echo "MISSING ARTIFACT: ${r}.json"; exit 1; }
+          done
+          for e in leaves.tsv traces.tsv causal_evidence.tsv collapse_feature_rows.tsv MANIFEST.tsv claims.tsv; do
+            test -s "ci_out/evidence/${e}" || { echo "MISSING EVIDENCE: ${e}"; exit 1; }
+          done
+
+      # Independent attestation (findings 2, 3): a SEPARATE process recomputes
+      # the Merkle root from the retained leaves and checks it against the
+      # producer's claim. It never reads compliance_report.json. A producer that
+      # reported a root inconsistent with its own leaves fails here.
+      - name: Independent attestation of evidence
+        run: cargo run --release -p attestation --bin attest -- ci_out
+
+      - name: Upload acceptance evidence
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: magicka-acceptance-evidence
+          path: ci_out/

.github/workflows/web-gates.yml

@@ -0,0 +1,72 @@
+name: magicka-web-gates
+
+# Phase H of plan2.md — the web CI gates.
+#
+# MERGE-BLOCKING gates live in merge-gates.yml (the merge_group-required job),
+# which runs the protocol + socket + replay/visibility/resilience suite. This
+# workflow provides the same Rust gates as fast PR/push feedback, plus the
+# rendered-browser layer.
+#
+# The Rust gates (web-rust-gates) enforce, deterministically and without a
+# browser:
+#   * 1,000 simulated matches, 0 replay hash mismatches  (determinism.rs)
+#   * 10,000 protocol fuzz cases, 0 server panics        (fuzz.rs)
+#   * 100 end-to-end matches over real sockets           (e2e.rs, protocol-level)
+#   * 0 hidden-state leaks                                (visibility.rs)
+#   * disconnect/reconnect + timer edges                 (resilience.rs)
+#
+# The rendered-browser layer (rendered-browser-e2e) is EXTERNAL-BLOCKED: it
+# cannot be merge-blocking until CI infrastructure with a real browser exists.
+# Until then it runs advisory-only (continue-on-error) and uploads Playwright
+# artifacts. It is NOT counted as satisfied coverage.
+
+on:
+  pull_request:
+    branches: [main]
+  push:
+    branches: [main]
+  workflow_dispatch:
+
+jobs:
+  web-rust-gates:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+      - name: Protocol / server / game-runtime unit tests
+        run: cargo test --release -p protocol -p game_runtime -p server -p web_assets -p web_client
+      - name: Web CI gates (1k matches, 10k fuzz, 100 e2e, leak + resilience)
+        run: cargo test --release -p web_tests
+
+  # EXTERNAL-BLOCKED: rendered-browser end-to-end. A real browser is not
+  # available in this CI, so this job is advisory only and never blocks merge.
+  # It produces Playwright artifacts as evidence; it does not satisfy the
+  # "rendered browser" coverage claim until CI infrastructure exists.
+  rendered-browser-e2e:
+    name: rendered-browser-e2e (ADVISORY — blocked on CI infra)
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    continue-on-error: true
+    needs: web-rust-gates
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "20"
+      - name: Install Playwright (Chromium)
+        working-directory: crates/web_tests/e2e
+        continue-on-error: true
+        run: |
+          npm install
+          npx playwright install --with-deps chromium
+      - name: Rendered-browser E2E (advisory)
+        working-directory: crates/web_tests/e2e
+        continue-on-error: true
+        run: npm test
+      - name: Upload advisory Playwright report
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: playwright-report-advisory
+          path: crates/web_tests/e2e/playwright-report
+          if-no-files-found: ignore

.gitignore

@@ -1,3 +1,9 @@
 /target
 /ci_out
+/ci_out_merge
 Cargo.lock
+# Playwright / Node artifacts for the rendered-browser E2E
+crates/web_tests/e2e/node_modules
+crates/web_tests/e2e/package-lock.json
+crates/web_tests/e2e/test-results
+crates/web_tests/e2e/playwright-report

Cargo.toml

@@ -11,6 +11,13 @@ members = [
     "crates/semantic_mutation",
     "crates/replay_corpus",
     "crates/ci_reports",
+    "crates/attestation",
+    "crates/protocol",
+    "crates/game_runtime",
+    "crates/web_assets",
+    "crates/web_client",
+    "crates/server",
+    "crates/web_tests",
 ]
 
 [workspace.package]

README.md

@@ -9,6 +9,43 @@ content, templates, or cosmetic runes — the value is in the tests that refuse
 to let the universe collapse into a single score, resource, effect axis,
 executor, rune, hidden formula, or decorative domain.
 
+## Compliance model
+
+No gate may pass from configuration, naming, shared implementation, smoke-scale
+runs, regenerated expectations, proxy metrics, a default profile, or a
+locally-runnable binary. A gate passes only from persisted, independently
+reproducible, full-scale adversarial evidence enforced at merge. Every
+acceptance obligation has all four of: a **measured artifact**, a **provenance
+chain** to the run that produced it, a **merge-blocking enforcement point**, and
+a **failure condition if the artifact or provenance is absent**.
+
+- The merge-blocking enforcement point is `.github/workflows/merge-gates.yml`,
+  whose `merge-gates` job runs `MAGICKA_PROFILE=merge` (full gates) and must be a
+  **required status check** on the protected branch / merge queue. It is not a
+  local binary, and the fast profile is advisory only — it can never stand in for
+  acceptance.
+- `compliance_report.json` enumerates every obligation, its artifact, its floor,
+  the actual measured value, and whether the artifact is present. A missing
+  required report fails acceptance.
+- The merge floors (50k worlds, 250k programs, 1,000,000 executions, 10
+  perturbations/exec, 100% reference/runtime comparison over base **and**
+  perturbations, 500 mutants, 10,000 replay cases) cannot be lowered by
+  environment overrides: a lowering override is recorded as a provenance failure
+  and the floor is kept.
+
+Every gate is built to be *able to fail*, and a negative-control test proves it does:
+
+| Gate | How it is made unbypassable | Negative control proving it can fail |
+|------|-----------------------------|--------------------------------------|
+| runtime_equivalence | Compares two **independent implementations** (the reference engine vs. `runtime_under_test::native`, which never calls the reference engine) | `buggy_runtime_is_rejected` — a runtime with one dropped causal edge is caught |
+| compression_resistance | Attacks operate on the **real serialized trace** (causal influence, info-flow, access, temporal, deltas), not a hash proxy; info loss is genuine unexplained variance | `single_factor_corpus_is_rejected` — a rank-1 universe is rejected |
+| mutation_survivor | Each mutant must fail the **named gate** it targets, not merely differ from the reference | `reference_passes_every_named_gate` + `no_mutant_survives_its_named_gate` |
+| replay | Expectations are **loaded from a committed file**, not regenerated in the same run | `corrupted_expectation_is_detected` |
+| domain_participation | Decorative/redundant domains are flagged directly | `decorative_domain_is_rejected` |
+| merge scale floor | Env overrides may only **raise** merge counts; a lowering attempt is recorded and the floor kept; executions actually performed are counted | `merge_floor_cannot_be_lowered_by_override`, `merge_profile_at_smoke_scale_is_rejected` |
+| 100% comparison | Reference vs. runtime-under-test compared for **every** execution — base and all perturbations, never base only | `runtime_equivalence` gate fails unless `equiv_total == base + perturbations` |
+| provenance | A Merkle root over per-execution records, plus independent engine identities, binds reported numbers to executed work | `merkle_root_binds_to_leaves` |
+
 ## Workspace layout
 
 Built in the mandatory order from the spec:
@@ -19,17 +56,18 @@ Built in the mandatory order from the spec:
 | – | `rune_ir` | Rune token / program model (no stream is ever rejected) |
 | 2 | `trace_model` | Execution trace + all graphs, behavior fingerprint, replay record, fault log, trace metrics |
 | 3 | `generators` | Worlds, programs, executors, contracts, perturbations; rejects flat cases |
-| 4 | `collapse_analysis` | The 11 compression attacks + collapse gates |
+| 4 | `collapse_analysis` | The 11 compression attacks over real trace structure + collapse gates |
-| 5 | `semantic_mutation` | Structurally generated mutant runtimes; proves every one is killed |
+| 5 | `semantic_mutation` | Structurally generated mutant runtimes; proves every one fails its named gate |
-| 6 | `replay_corpus` | Permanent, bit-exact replay cases |
+| 6 | `replay_corpus` | Permanent, bit-exact replay cases persisted to `corpus/replay_corpus.tsv` |
 | 7 | `reference_runtime` | The executable spec engine (`Runtime` trait, `resolve`) |
-| 8 | `runtime_under_test` | Config-driven engine, proven equivalent to the reference |
+| 8 | `runtime_under_test` | An **independent** interpreter (`native`) proven equivalent to the reference |
-| – | `ci_reports` | Orchestrator + `ci` binary; emits the 8 required reports |
+| – | `ci_reports` | Orchestrator + `ci` binary; emits 8 gate reports + a provenance report |
 
-The optimized runtime may not begin before steps 1–7 pass CI; until then the
+The runtime under test does not call the reference engine. It re-derives the
-runtime under test is the reference engine driven through the same config
+canonical behavior from the spec in a different code organization, so 100%
-surface, which is equivalent by construction. Mutation swaps in a *mutated*
+agreement is *evidence* the spec is implemented correctly rather than a
-config to prove the suite detects any divergence.
+tautology. (`native_matches_reference_bit_for_bit` checks this over a 2000-seed
+sweep.)
 
 ## The engine in one paragraph
 
@@ -40,53 +78,149 @@ domains, mixes them through a nonlinear avalanche keyed by per-domain
 constants, the world coupling, and the executor's salt, then writes back —
 recording causal/read/write/information-flow/temporal edges as it goes.
 Scheduled effects and coupling diffusion propagate changes 3 turns into the
-future. Because every output bit depends on all inputs, the universe is
+future.
-high-rank, incompressible, future-sensitive, and executor-divergent — exactly
-the properties the gates demand.
 
 ## Running CI
 
 ```bash
-cargo test                                   # unit tests for every crate
+cargo test                                          # unit tests + negative controls
-cargo run --release -p ci_reports --bin ci   # full gate run (fast scale)
+MAGICKA_PROFILE=fast  cargo run --release -p ci_reports --bin ci   # advisory PR slice
+MAGICKA_PROFILE=merge cargo run --release -p ci_reports --bin ci   # acceptance (full gates)
 ```
 
-Reports are written to `ci_out/` (8 JSON files + `ci_summary.md`). The binary
+Reports are written to the output dir (8 gate reports + `provenance_report.json`
-exits non-zero if any gate fails.
++ `compliance_report.json` + `ci_summary.md`). The binary exits non-zero if any
+gate fails or any required artifact is absent.
 
-### Scales
+### Profiles
 
-`MAGICKA_SCALE` selects the corpus size; individual counts can be overridden.
+`MAGICKA_PROFILE` (or `MAGICKA_SCALE`) selects the run profile.
 
-| Scale | executions | replay | mutants | notes |
+| Profile | executions | replay | mutants | role |
-|-------|-----------|--------|---------|-------|
+|---------|-----------|--------|---------|------|
-| `tiny` | 120 | 120 | 520 | smoke (~80 ms) |
+| `fast` (default) | 600 | 10,000 (committed) | 520 | **advisory only — never acceptance** |
-| `fast` (default) | 600 | 600 | 520 | every gate, ~0.3 s |
+| `tiny` | 120 | 10,000 | 520 | smoke |
-| `full` | 1,000,000 | 10,000 | 600 | merge-blocking spec gates |
+| `merge` (`MAGICKA_SCALE=full`) | 1,000,000 | 10,000 | 600 | **acceptance — hard floors** |
+
+The fast/tiny profiles print `ADVISORY … NOT a merge-blocking acceptance run`
+and are labelled non-acceptance in `compliance_report.json`. Acceptance comes
+only from the merge profile, run by the merge-gates workflow. The merge floors
+cannot be lowered by environment overrides (a lowering override is recorded as a
+provenance failure and the floor kept).
+
+### Merge-blocking enforcement (required check)
+
+`.github/workflows/merge-gates.yml` defines the enforcement point. Configure
+branch protection / the merge queue to **require** the `merge-gates` job. That
+job runs the full merge profile, verifies the committed corpus has ≥10,000
+cases, and fails if any required artifact is missing. The full run executes
+~1M base executions × (1 base + 10 perturbations) with 100% reference/runtime
+comparison; it completes in minutes on a CI runner.
+
+### Replay corpus
+
+The replay corpus is committed at
+`crates/replay_corpus/corpus/replay_corpus.tsv` (10,000 cases). Replay loads
+those expectations and re-executes the reference, so any engine change that
+alters a hash makes the committed file and the fresh run disagree and CI fails.
+Regenerate it only as a deliberate, reviewed migration:
 
 ```bash
-MAGICKA_SCALE=full cargo run --release -p ci_reports --bin ci
+cargo run --release -p replay_corpus --bin freeze -- 10000
-# or override individual counts:
-MAGICKA_EXECUTIONS=20000 MAGICKA_REPLAY=10000 cargo run --release -p ci_reports --bin ci
 ```
 
-The gate *thresholds* are identical across scales — only the corpus size
-changes. Fast CI runs a representative slice; merge-blocking CI runs `full`.
-
-## Gates enforced (all must pass)
-
-- **runtime_equivalence** — 100% of executions: `canonical(reference) == canonical(runtime_under_test)` over delta, trace, faults, replay hash, and 3-turn future hash.
-- **causal_rank / trace** — median causal edges ≥ 24, 95% causal rank ≥ 6, median touched domains ≥ 4, 95% ≥ 3, fingerprint collision rate < 5%, largest cluster < 2%.
-- **domain_participation** — each domain appears in ≥ 35% of traces, influences ≥ 20%, is mutated in ≥ 20%; removing any domain loses ≥ 10% behavioral diversity; merging any pair loses ≥ 8%; no read-only or write-only domain.
-- **metamorphic_response** — ≥ 90% of perturbations alter the trace, ≥ 75% the delta, ≥ 50% the 3-turn future; ≤ 5% unexplained neutral.
-- **compression_resistance** — best 1/2/4-factor models predict < 40/55/70%; no single domain > 30%, no pair > 55%; every compressed model loses ≥ 35% information. All 11 attack families are run.
-- **mutation_survivor** — ≥ 500 structurally generated mutants, 0 survivors.
-- **contract** — every admitted case satisfies its semantic contract (min causal rank, domain participation, future sensitivity, context divergence, max compressibility); contract-violating cases are rejected at admission.
-- **replay** — 100% deterministic, 0 hash drift.
-- **coverage** — generated/contract rejection accounting; no admitted case fails the generated gates.
-
 ## Determinism
 
 Everything is seed-derived and integer-only (SplitMix64 RNG, FNV-1a content
 hashing, wrapping/guarded arithmetic). No floating point enters a canonical
 hash, so replay is bit-exact across machines and runs. No external crates.
+
+## The web game (plan2.md)
+
+A browser game is built **around** the existing runtime — it is a playable
+window into the Rust universe, never a second simulation. The browser sends only
+*intent*; the server is the sole authority; every rune program executes through
+the **independent** interpreter (`runtime_under_test::native_resolve`) against
+the shared world. The game deliberately does **not** call the reference engine —
+the interpreter it uses is the one the runtime-equivalence gate proves correct
+(with a negative control proving that gate can fail). Same constraints as the
+rest of the repo: pure `std`, no external crates (the WebSocket server
+hand-rolls SHA-1, base64, and RFC 6455 framing; JSON is hand-rolled with a total
+parser).
+
+> Audit note: the hand-rolled SHA-1 / base64 / RFC-6455 framing and JSON parser
+> are checked against published test vectors (RFC 6455 §1.3 accept key, SHA-1
+> "abc", base64 length cases) and a fuzz gate, but they are bespoke
+> cryptographic/parsing code and carry audit risk relative to a reviewed
+> library. They exist to honor the repo's no-external-crates rule; a future
+> hardening pass could swap in vetted implementations behind the same interface.
+
+```
+Rust runtime → game_runtime (authority) → protocol (WS messages) → server → browser
+```
+
+| Crate | Role |
+|-------|------|
+| `protocol` | Versioned, hashable, **total-decode** client/server messages + JSON value/parser. A malformed packet yields `Err`, never a panic. |
+| `game_runtime` | Authoritative match state. Resolves turns through the **independent interpreter** (`runtime_under_test`, not the reference engine), filters visibility/knowledge, records + regenerates replays. A match is a pure function of `(seed, roster, ordered inputs)`. |
+| `web_assets` | The embedded browser client (HTML/CSS/JS): arena, rune editor, domain/knowledge panels, replay viewer. |
+| `web_client` | Static-asset HTTP delivery (keeps raw assets separate from framing). |
+| `server` | `std::net` HTTP + WebSocket server: turn timer, action collection, disconnect handling, panic-proof dispatch. |
+| `web_tests` | A dependency-free WebSocket test client + the Phase H gates. |
+
+### Running it
+
+```bash
+cargo run --release -p server --bin magicka-server   # serve on 127.0.0.1:8080
+# then open http://127.0.0.1:8080 in a browser
+MAGICKA_ADDR=0.0.0.0:9000 MAGICKA_TURN_MS=8000 cargo run --release -p server --bin magicka-server
+```
+
+Join is immediate (1 player + a training dummy). A duel shares a match by id:
+two browsers that `JoinMatch` the same `match_id` take slots 1 and 2.
+
+### Web CI gates (Phase H)
+
+These gates are **merge-blocking**: they run inside the merge-required job in
+`.github/workflows/merge-gates.yml` (and as fast PR feedback in
+`web-gates.yml`). They are the Rust suite in `crates/web_tests`, run with
+`cargo test -p web_tests`:
+
+| Gate | Test | Minimum | Status |
+|------|------|---------|--------|
+| Replay determinism | `determinism.rs` | 1,000 simulated matches, **0 hash mismatches** | merge-blocking |
+| Protocol fuzz | `fuzz.rs` | 10,000 fuzz cases, **0 panics** (+ a live server survives a malformed-packet burst) | merge-blocking |
+| End-to-end matches | `e2e.rs` | **100** full matches over real sockets; recorded replay reproduces every live per-turn hash | merge-blocking |
+| Hidden-state leaks | `visibility.rs` | **0 leaks** — no client-bound frame carries a hidden key; redaction counts every withheld value | merge-blocking |
+| Disconnect / timer edges | `resilience.rs` | mid-match disconnect does not corrupt the match; wrong-turn / late submits are rejected deterministically | merge-blocking |
+| Rendered-browser E2E | `e2e/specs/play.spec.js` | a real browser joins, casts, and replays a match | **external-blocked (advisory only)** |
+
+Scope honesty — two distinct things, not conflated:
+
+- The "100 E2E matches" merge-blocking gate drives the full
+  HTTP→WebSocket→protocol→runtime path **headlessly over real sockets**. This is
+  protocol-level coverage. It is **not** rendered-browser coverage and is not
+  claimed as such.
+- Rendered-browser coverage is **blocked on CI infrastructure**: this CI has no
+  real browser, so the Playwright suite under `crates/web_tests/e2e/` cannot be
+  merge-blocking yet. It runs **advisory-only** (`continue-on-error`) in the
+  `rendered-browser-e2e` job and uploads its report as an artifact. Until a CI
+  runner with a browser exists, rendered-browser E2E is treated as
+  **unsatisfied**, not green. Run it locally with:
+
+```bash
+cd crates/web_tests/e2e && npm install && npx playwright install chromium && npm test
+```
+
+### Acceptance criteria mapping (plan2.md)
+
+| Criterion | Where it holds |
+|-----------|----------------|
+| A player can join a browser match | `server` join + `web_assets` client; `e2e.rs::single_match_full_playthrough` |
+| A turn timer runs | `server` timer thread; client header countdown |
+| Inspect / move / attack / cast | `Action` in `protocol`; `game_runtime::apply_action` |
+| Rune programs execute only on the server | `game_runtime` is the only caller of the interpreter (`runtime_under_test::native_resolve`); client never imports `EngineConfig` (asserted in `web_assets`) |
+| Results return as filtered observations | `VisibleWorldSnapshot`; `visibility.rs` |
+| Replay can reproduce the match | `game_runtime::replay`; `determinism.rs`, `e2e.rs` |
+| Browser cannot alter hidden truth | intent-only protocol; `visibility.rs` leak gate |
+| CI proves protocol, replay, visibility, authority | merge-blocking gates in `merge-gates.yml` (+ `web-gates.yml`); rendered-browser E2E remains external-blocked |

crates/attestation/Cargo.toml

@@ -0,0 +1,22 @@
+[package]
+name = "attestation"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[[bin]]
+name = "attest"
+path = "src/main.rs"
+
+[dependencies]
+# NOT ci_reports. The root, the full-trace hashes, and the leaves are recomputed
+# by this crate's own code path — an independent check, not a re-export of the
+# producer's claim. trace_model/world_model provide the shared trace/delta types
+# and hashing the verifier reconstructs from raw evidence.
+world_model = { path = "../world_model" }
+trace_model = { path = "../trace_model" }
+# For independent RE-EXECUTION of retained per-edge causal interventions.
+generators = { path = "../generators" }
+reference_runtime = { path = "../reference_runtime" }
+# For recomputing collapse feature rows from full traces (proving derivation).
+collapse_analysis = { path = "../collapse_analysis" }

crates/attestation/src/lib.rs

@@ -0,0 +1,596 @@
+//! `attestation` — an **independent** verifier of a CI run's evidence
+//! (findings 2 and 3).
+//!
+//! The substitution it removes: a `compliance_report.json` written by the same
+//! binary that ran the gates is self-certification, and a Merkle *root* reported
+//! without its *leaves* cannot be checked by anyone. This crate is a separate
+//! process with its own code path that:
+//!   * reads the retained per-execution leaves (`evidence/leaves.tsv`),
+//!   * recomputes the Merkle root from them with its own implementation,
+//!   * reads the producer's *claims* (`evidence/claims.tsv`) and checks the
+//!     recomputed root and leaf count match what was claimed,
+//!   * NEVER reads `compliance_report.json` — it does not trust the producer's
+//!     own pass/fail verdict.
+//!
+//! It depends only on `world_model` for the shared hash primitive. What it does
+//! NOT do (and cannot, in-repo) is prove the leaves correspond to real
+//! executions performed by a trusted third party — that requires external
+//! re-execution / signing infrastructure (see the BLOCKED note in the report).
+
+use generators::generate_accepted_case;
+use reference_runtime::{execute, EngineConfig, ResolutionInput};
+use std::collections::HashSet;
+use std::path::Path;
+use trace_model::{ExecutionTrace, ReplayRecord};
+use world_model::{Hash, Hasher, WorldDelta, HIDDEN_LANES, LANES, NUM_DOMAINS};
+
+/// Re-execute one retained causal record from its seed and recompute the
+/// destination delta before and after perturbing the recorded source lane.
+/// Returns `(base_dv, alt_dv)`; the caller compares against the retained values.
+pub fn recompute_causal_record(
+    seed: u64,
+    from_domain: usize,
+    from_lane: usize,
+    from_hidden: bool,
+    to_domain: usize,
+    to_lane: usize,
+    to_hidden: bool,
+) -> (i64, i64) {
+    let (case, _) = generate_accepted_case(seed);
+    let cfg = EngineConfig::reference();
+    let input = ResolutionInput {
+        world: case.world.clone(),
+        program: case.program.clone(),
+        contexts: case.contexts.clone(),
+        contract_seed: case.contract_seed,
+        perturbation_seed: case.perturbation_seed,
+    };
+    let base = execute(&cfg, &input);
+    let mut w = input.world.clone();
+    let fd = from_domain % NUM_DOMAINS;
+    if from_hidden {
+        let l = from_lane % HIDDEN_LANES;
+        w.domains[fd].hidden[l] = w.domains[fd].hidden[l].wrapping_add(0x9_27c1);
+    } else {
+        let l = from_lane % LANES;
+        w.domains[fd].observed[l] = w.domains[fd].observed[l].wrapping_add(0x9_27c1);
+    }
+    let mut alt_input = input.clone();
+    alt_input.world = w;
+    let alt = execute(&cfg, &alt_input);
+    let dd = to_domain % NUM_DOMAINS;
+    let read = |d: &world_model::DomainDelta| -> i64 {
+        if to_hidden {
+            d.hidden[to_lane % HIDDEN_LANES]
+        } else {
+            d.observed[to_lane % LANES]
+        }
+    };
+    (read(&base.delta.domain_deltas[dd]), read(&alt.delta.domain_deltas[dd]))
+}
+
+/// Recompute the Merkle root over `leaves` (binary tree, FNV-combined). This is
+/// an independent re-implementation of the producer's algorithm; agreement is
+/// the check.
+pub fn merkle_root(leaves: &[u64]) -> u64 {
+    if leaves.is_empty() {
+        return 0;
+    }
+    let mut level = leaves.to_vec();
+    while level.len() > 1 {
+        let mut next = Vec::with_capacity(level.len().div_ceil(2));
+        for pair in level.chunks(2) {
+            let mut h = Hasher::new();
+            h.write_tag("merkle");
+            h.write_u64(pair[0]);
+            h.write_u64(if pair.len() > 1 { pair[1] } else { pair[0] });
+            next.push(h.finish().0);
+        }
+        level = next;
+    }
+    level[0]
+}
+
+/// Parse a leaves file: one hex u64 per data line, `#`/`leaf` headers skipped.
+pub fn parse_leaves(text: &str) -> Vec<u64> {
+    text.lines()
+        .filter(|l| !l.starts_with('#') && !l.starts_with("leaf") && !l.trim().is_empty())
+        .filter_map(|l| u64::from_str_radix(l.trim(), 16).ok())
+        .collect()
+}
+
+/// Parse a `key<TAB>value` claims file into `(key, value)` pairs.
+pub fn parse_claims(text: &str) -> Vec<(String, String)> {
+    text.lines()
+        .filter(|l| !l.starts_with('#') && !l.trim().is_empty())
+        .filter_map(|l| l.split_once('\t').map(|(k, v)| (k.trim().to_string(), v.trim().to_string())))
+        .collect()
+}
+
+fn claim<'a>(claims: &'a [(String, String)], key: &str) -> Option<&'a str> {
+    claims.iter().find(|(k, _)| k == key).map(|(_, v)| v.as_str())
+}
+
+/// Reconstruct the leaf of one full-trace evidence record from its raw trace +
+/// delta + seeds, independently of any reported digest. Returns the recomputed
+/// leaf, or `None` if the record is malformed.
+///
+/// `line` is `kind ws ps cs prs future leaf <TAB> <trace> <TAB> <delta>` where
+/// `kind` is `b` (base) or `p` (perturbation). Returns `(recomputed_leaf,
+/// claimed_leaf, is_base)`.
+pub fn recompute_trace_leaf(line: &str) -> Option<(u64, u64, bool)> {
+    let mut parts = line.splitn(3, '\t');
+    let header = parts.next()?;
+    let trace_s = parts.next()?;
+    let delta_s = parts.next()?;
+    let mut h = header.split_whitespace();
+    let kind = h.next()?;
+    let is_base = kind == "b";
+    let hx = |s: Option<&str>| -> Option<u64> { u64::from_str_radix(s?.trim(), 16).ok() };
+    let world_seed = hx(h.next())?;
+    let program_seed = hx(h.next())?;
+    let contract_seed = hx(h.next())?;
+    let perturbation_seed = hx(h.next())?;
+    let future = hx(h.next())?;
+    let claimed_leaf = hx(h.next())?;
+
+    let trace = ExecutionTrace::deserialize(trace_s)?;
+    let delta = WorldDelta::deserialize(delta_s)?;
+    // Recompute the canonical trace hash and delta hash from the FULL structure.
+    let rr = ReplayRecord {
+        world_seed,
+        program_seed,
+        contract_seed,
+        perturbation_seed,
+        trace_hash: trace.canonical_hash(),
+        delta_hash: delta.hash(),
+        future_hash: Hash(future),
+    };
+    Some((rr.hash().0, claimed_leaf, is_base))
+}
+
+/// Outcome of an attestation.
+#[derive(Clone, Debug)]
+pub struct Attestation {
+    pub ok: bool,
+    pub leaf_count: usize,
+    pub recomputed_root: u64,
+    pub claimed_root: Option<u64>,
+    pub traces_verified: usize,
+    pub traces_total: usize,
+    pub causal_total: usize,
+    pub causal_recomputed: usize,
+    pub causal_confirmed: usize,
+    pub collapse_total: usize,
+    pub collapse_derived: usize,
+    pub checks: Vec<(String, bool)>,
+}
+
+/// Verify the evidence directory `dir` (which contains `evidence/`).
+pub fn verify_dir(dir: &Path) -> Result<Attestation, String> {
+    let leaves_path = dir.join("evidence/leaves.tsv");
+    let claims_path = dir.join("evidence/claims.tsv");
+    let leaves_txt = std::fs::read_to_string(&leaves_path)
+        .map_err(|e| format!("cannot read {}: {e}", leaves_path.display()))?;
+    let claims_txt = std::fs::read_to_string(&claims_path)
+        .map_err(|e| format!("cannot read {}: {e}", claims_path.display()))?;
+
+    let leaves = parse_leaves(&leaves_txt);
+    let claims = parse_claims(&claims_txt);
+    let recomputed = merkle_root(&leaves);
+
+    let claimed_root = claim(&claims, "root").and_then(|v| u64::from_str_radix(v, 16).ok());
+    let claimed_count = claim(&claims, "leaf_count").and_then(|v| v.parse::<usize>().ok());
+    let claimed_comparisons =
+        claim(&claims, "total_comparisons").and_then(|v| v.parse::<usize>().ok());
+
+    let mut checks = Vec::new();
+    let root_ok = claimed_root == Some(recomputed);
+    checks.push(("recomputed_root == claimed_root".into(), root_ok));
+    let count_ok = claimed_count == Some(leaves.len());
+    checks.push(("leaf_count == claimed_leaf_count".into(), count_ok));
+    let cmp_ok = claimed_comparisons.map(|c| c == leaves.len()).unwrap_or(false);
+    checks.push(("leaf_count == claimed_total_comparisons".into(), cmp_ok));
+    let nonempty = !leaves.is_empty();
+    checks.push(("leaves are present (root has leaves)".into(), nonempty));
+
+    // Full-trace evidence (finding 4): reconstruct each sampled trace + delta,
+    // recompute its leaf independently, and confirm it both matches the record's
+    // claimed leaf AND is one of the retained leaves the root is built from.
+    let leaf_set: HashSet<u64> = leaves.iter().copied().collect();
+    let traces_path = dir.join("evidence/traces.tsv");
+    let traces_txt = std::fs::read_to_string(&traces_path)
+        .map_err(|e| format!("cannot read {}: {e}", traces_path.display()))?;
+    let mut traces_total = 0usize;
+    let mut traces_verified = 0usize;
+    let mut covered: HashSet<u64> = HashSet::new();
+    for line in traces_txt.lines() {
+        if line.starts_with('#') || line.trim().is_empty() {
+            continue;
+        }
+        traces_total += 1;
+        if let Some((recomputed_leaf, claimed_leaf, _is_base)) = recompute_trace_leaf(line) {
+            if recomputed_leaf == claimed_leaf && leaf_set.contains(&recomputed_leaf) {
+                traces_verified += 1;
+                covered.insert(recomputed_leaf);
+            }
+        }
+    }
+    let traces_present = traces_total > 0;
+    checks.push(("full-trace evidence present".into(), traces_present));
+    checks.push((
+        "every full trace recomputes to a retained leaf".into(),
+        traces_present && traces_verified == traces_total,
+    ));
+    // Completeness (finding 3): every retained leaf must be covered by a full
+    // trace record, and there must be exactly one record per leaf — a bijection,
+    // not just a consistent count.
+    checks.push((
+        "every leaf has full retained evidence (bijection)".into(),
+        traces_present
+            && traces_total == leaves.len()
+            && covered.len() == leaf_set.len(),
+    ));
+
+    // Causal intervention evidence (finding 7): RE-EXECUTE each retained record
+    // from its seed and confirm the retained base_dv/alt_dv reproduce, then
+    // require the confirmed fraction to clear the threshold.
+    let causal_path = dir.join("evidence/causal_evidence.tsv");
+    let causal_txt = std::fs::read_to_string(&causal_path)
+        .map_err(|e| format!("cannot read {}: {e}", causal_path.display()))?;
+    let mut causal_total = 0usize;
+    let mut causal_recomputed = 0usize;
+    let mut causal_confirmed = 0usize;
+    for line in causal_txt.lines() {
+        if line.starts_with("seed") || line.starts_with('#') || line.trim().is_empty() {
+            continue;
+        }
+        let f: Vec<&str> = line.split('\t').collect();
+        if f.len() < 9 {
+            continue;
+        }
+        let seed = match u64::from_str_radix(f[0].trim(), 16) {
+            Ok(v) => v,
+            Err(_) => continue,
+        };
+        let p = |i: usize| f[i].trim().parse::<i64>().ok();
+        let (Some(fd), Some(fl), Some(fh), Some(td), Some(tl), Some(th), Some(bdv), Some(adv)) =
+            (p(1), p(2), p(3), p(4), p(5), p(6), p(7), p(8))
+        else {
+            continue;
+        };
+        causal_total += 1;
+        let (rb, ra) = recompute_causal_record(
+            seed, fd as usize, fl as usize, fh != 0, td as usize, tl as usize, th != 0,
+        );
+        if rb == bdv && ra == adv {
+            causal_recomputed += 1;
+        }
+        if rb != ra {
+            causal_confirmed += 1;
+        }
+    }
+    // Collapse derivation (finding 4): recompute each collapse feature row from
+    // the corresponding retained FULL trace + delta and require bit-exact match,
+    // proving the summary the collapse gate consumed derives from the full trace.
+    let crows_path = dir.join("evidence/collapse_feature_rows.tsv");
+    let crows_txt = std::fs::read_to_string(&crows_path)
+        .map_err(|e| format!("cannot read {}: {e}", crows_path.display()))?;
+    let claimed_rows: Vec<Vec<u64>> = crows_txt
+        .lines()
+        .filter(|l| !l.trim().is_empty())
+        .map(|l| l.split_whitespace().filter_map(|t| t.parse::<u64>().ok()).collect())
+        .collect();
+    // Re-read the trace records in order to recompute their feature rows.
+    let mut collapse_total = 0usize;
+    let mut collapse_derived = 0usize;
+    {
+        let mut idx = 0usize;
+        for line in traces_txt.lines() {
+            // Collapse rows correspond to BASE executions in order.
+            if !line.starts_with("b ") {
+                continue;
+            }
+            if idx >= claimed_rows.len() {
+                break;
+            }
+            let mut parts = line.splitn(3, '\t');
+            let _hdr = parts.next();
+            let trace_s = parts.next();
+            let delta_s = parts.next();
+            if let (Some(ts), Some(ds)) = (trace_s, delta_s) {
+                if let (Some(trace), Some(delta)) =
+                    (ExecutionTrace::deserialize(ts), WorldDelta::deserialize(ds))
+                {
+                    let recomputed = collapse_analysis::trace_feature_row(&trace, &delta);
+                    let recomputed_bits: Vec<u64> = recomputed.iter().map(|v| v.to_bits()).collect();
+                    collapse_total += 1;
+                    if recomputed_bits == claimed_rows[idx] {
+                        collapse_derived += 1;
+                    }
+                }
+            }
+            idx += 1;
+        }
+    }
+    let collapse_present = collapse_total > 0;
+    checks.push(("collapse feature rows present".into(), collapse_present));
+    checks.push((
+        "every collapse summary derives from a full trace".into(),
+        collapse_present && collapse_derived == collapse_total,
+    ));
+
+    let causal_present = causal_total > 0;
+    let causal_frac = if causal_total > 0 { causal_confirmed as f64 / causal_total as f64 } else { 0.0 };
+    checks.push(("causal evidence present".into(), causal_present));
+    checks.push((
+        "every causal record recomputes (base_dv/alt_dv reproduce)".into(),
+        causal_present && causal_recomputed == causal_total,
+    ));
+    checks.push((
+        "recomputed causal confirmation >= 0.50".into(),
+        causal_present && causal_frac >= 0.50,
+    ));
+
+    // Artifact bundle (finding 1): the merge-scale claim requires a complete,
+    // content-hashed bundle. Verify the manifest lists the required files and
+    // each file's recomputed hash + length match.
+    let manifest_path = dir.join("evidence/MANIFEST.tsv");
+    let manifest_txt = std::fs::read_to_string(&manifest_path)
+        .map_err(|e| format!("cannot read {}: {e}", manifest_path.display()))?;
+    let required = [
+        "leaves.tsv",
+        "claims.tsv",
+        "causal_evidence.tsv",
+        "collapse_feature_rows.tsv",
+        "traces.tsv",
+    ];
+    let mut listed: HashSet<String> = HashSet::new();
+    let mut bundle_intact = true;
+    for line in manifest_txt.lines() {
+        if line.starts_with("file") || line.trim().is_empty() {
+            continue;
+        }
+        let f: Vec<&str> = line.split('\t').collect();
+        if f.len() < 3 {
+            bundle_intact = false;
+            continue;
+        }
+        let name = f[0].trim();
+        let claimed_hash = u64::from_str_radix(f[1].trim(), 16).ok();
+        let claimed_len = f[2].trim().parse::<usize>().ok();
+        let bytes = std::fs::read(dir.join("evidence").join(name)).unwrap_or_default();
+        let mut h = Hasher::new();
+        h.write_tag("evidence-file");
+        h.write_bytes(&bytes);
+        if claimed_hash != Some(h.finish().0) || claimed_len != Some(bytes.len()) {
+            bundle_intact = false;
+        }
+        listed.insert(name.to_string());
+    }
+    let bundle_complete = required.iter().all(|r| listed.contains(*r));
+    checks.push(("artifact bundle manifest complete".into(), bundle_complete));
+    checks.push(("artifact bundle files intact (hash + length)".into(), bundle_intact && bundle_complete));
+
+    let ok = checks.iter().all(|(_, b)| *b);
+    Ok(Attestation {
+        ok,
+        leaf_count: leaves.len(),
+        recomputed_root: recomputed,
+        claimed_root,
+        traces_verified,
+        traces_total,
+        causal_total,
+        causal_recomputed,
+        causal_confirmed,
+        collapse_total,
+        collapse_derived,
+        checks,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn root_matches_known_vector() {
+        // Mirrors ci_reports::merkle_root over the same leaves.
+        let a = merkle_root(&[1, 2, 3]);
+        let b = merkle_root(&[1, 2, 3]);
+        let c = merkle_root(&[1, 2, 4]);
+        assert_eq!(a, b);
+        assert_ne!(a, c);
+    }
+
+    use trace_model::{
+        BehaviorFingerprint, CausalEdge, CausalGraph, CausalNode, DivergenceGraph,
+        DomainAccessGraph, ExecutionTrace, InformationFlowGraph, PerturbationResponse, ReplayRecord,
+        TemporalGraph,
+    };
+    use world_model::{DomainDelta, DomainId, Hash, WorldDelta, HIDDEN_LANES, LANES};
+
+    fn sample_trace(seed: u64) -> ExecutionTrace {
+        let mut read_graph = DomainAccessGraph::default();
+        read_graph.access_count[(seed % 8) as usize] = 3;
+        read_graph.edges.push((0, 2, (seed % 7) as u32 + 1));
+        let mut write_graph = DomainAccessGraph::default();
+        write_graph.access_count[2] = 4;
+        let causal_graph = CausalGraph {
+            edges: vec![CausalEdge {
+                from: CausalNode { domain: 0, lane: 1, hidden: false, step: 2 },
+                to: CausalNode { domain: 2, lane: 0, hidden: true, step: 2 },
+                weight: seed as i64 - 100,
+            }],
+        };
+        ExecutionTrace {
+            read_graph,
+            write_graph,
+            causal_graph,
+            information_flow: InformationFlowGraph { edges: vec![(0, 2, 9)] },
+            executor_divergence: DivergenceGraph { executor_count: 2, pairwise: vec![0.0, 0.5, 0.5, 0.0] },
+            temporal_graph: TemporalGraph { edges: vec![(1, 2, 3)] },
+            perturbation_response: PerturbationResponse::default(),
+            behavior_fingerprint: BehaviorFingerprint::from_features(vec![seed as i64, -2, 3]),
+        }
+    }
+
+    fn sample_delta(seed: u64) -> WorldDelta {
+        let domain_deltas = (0..NUM_DOMAINS)
+            .map(|d| {
+                let mut observed = [0i64; LANES];
+                observed[0] = seed as i64 + d as i64;
+                DomainDelta { domain: DomainId(d as u8), observed, hidden: [0i64; HIDDEN_LANES] }
+            })
+            .collect();
+        WorldDelta { domain_deltas, turn_advance: 0 }
+    }
+
+    fn record(seed: u64) -> (ReplayRecord, ExecutionTrace, WorldDelta) {
+        let trace = sample_trace(seed);
+        let delta = sample_delta(seed);
+        let rr = ReplayRecord {
+            world_seed: seed,
+            program_seed: seed ^ 1,
+            contract_seed: seed ^ 2,
+            perturbation_seed: seed ^ 3,
+            trace_hash: trace.canonical_hash(),
+            delta_hash: delta.hash(),
+            future_hash: Hash(seed.wrapping_mul(0x9e3779b97f4a7c15)),
+        };
+        (rr, trace, delta)
+    }
+
+    /// Write a full evidence dir from real records. If `tamper_trace` is set,
+    /// that record's serialized trace is corrupted after its leaf was claimed.
+    fn write_full_evidence(dir: &Path, seeds: &[u64], tamper_trace: Option<usize>) {
+        let ev = dir.join("evidence");
+        std::fs::create_dir_all(&ev).unwrap();
+        let records: Vec<_> = seeds.iter().map(|&s| record(s)).collect();
+        let leaves: Vec<u64> = records.iter().map(|(rr, _, _)| rr.hash().0).collect();
+
+        let mut lt = String::from("leaf\n");
+        for l in &leaves {
+            lt.push_str(&format!("{:016x}\n", l));
+        }
+        std::fs::write(ev.join("leaves.tsv"), lt).unwrap();
+
+        let claims = format!(
+            "# claims\nroot\t{:016x}\nleaf_count\t{}\ntotal_comparisons\t{}\n",
+            merkle_root(&leaves),
+            leaves.len(),
+            leaves.len()
+        );
+        std::fs::write(ev.join("claims.tsv"), claims).unwrap();
+
+        let mut traces = String::from("# trace evidence\n");
+        for (i, (rr, trace, delta)) in records.iter().enumerate() {
+            let mut trace_s = trace.serialize();
+            if tamper_trace == Some(i) {
+                // Corrupt the full trace without changing the claimed leaf.
+                trace_s = trace_s.replacen("trace-v1 ", "trace-v1 999 ", 1);
+            }
+            traces.push_str(&format!(
+                "b {:016x} {:016x} {:016x} {:016x} {:016x} {:016x}\t{}\t{}\n",
+                rr.world_seed, rr.program_seed, rr.contract_seed, rr.perturbation_seed,
+                rr.future_hash.0, rr.hash().0, trace_s, delta.serialize(),
+            ));
+        }
+        std::fs::write(ev.join("traces.tsv"), traces).unwrap();
+
+        // Real, recomputable causal evidence from actual reference executions.
+        let mut causal = String::from(
+            "seed\tfrom_domain\tfrom_lane\tfrom_hidden\tto_domain\tto_lane\tto_hidden\tbase_dv\talt_dv\n",
+        );
+        let cfg = EngineConfig::reference();
+        for &cseed in &[0xC0FFEEu64, 0xBEEF, 0x1234, 0x5EED, 0xABCD] {
+            let (case, _) = generate_accepted_case(cseed);
+            let input = ResolutionInput {
+                world: case.world.clone(),
+                program: case.program.clone(),
+                contexts: case.contexts.clone(),
+                contract_seed: case.contract_seed,
+                perturbation_seed: case.perturbation_seed,
+            };
+            let base = execute(&cfg, &input);
+            let edges = base.trace.causal_graph.edges.clone();
+            let stride = (edges.len() / 12).max(1);
+            for e in edges.iter().step_by(stride).take(12) {
+                let fd = e.from.domain as usize % NUM_DOMAINS;
+                let fl = e.from.lane as usize;
+                let fh = e.from.hidden;
+                let td = e.to.domain as usize % NUM_DOMAINS;
+                let tl = e.to.lane as usize;
+                let th = e.to.hidden;
+                let (b, a) = recompute_causal_record(cseed, fd, fl, fh, td, tl, th);
+                causal.push_str(&format!(
+                    "{:016x}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\n",
+                    cseed, fd, fl, fh as u8, td, tl, th as u8, b, a
+                ));
+            }
+        }
+        std::fs::write(ev.join("causal_evidence.tsv"), causal).unwrap();
+
+        // Collapse feature rows derived from the same synthetic traces, in order.
+        let mut crows = String::new();
+        for (_, trace, delta) in &records {
+            let row = collapse_analysis::trace_feature_row(trace, delta);
+            let cells: Vec<String> = row.iter().map(|v| v.to_bits().to_string()).collect();
+            crows.push_str(&cells.join(" "));
+            crows.push('\n');
+        }
+        std::fs::write(ev.join("collapse_feature_rows.tsv"), crows).unwrap();
+
+        // Bundle manifest over the written files.
+        let mut manifest = String::from("file\thash\tbytes\n");
+        for name in ["leaves.tsv", "claims.tsv", "causal_evidence.tsv", "collapse_feature_rows.tsv", "traces.tsv"] {
+            let bytes = std::fs::read(ev.join(name)).unwrap_or_default();
+            let mut h = Hasher::new();
+            h.write_tag("evidence-file");
+            h.write_bytes(&bytes);
+            manifest.push_str(&format!("{}\t{:016x}\t{}\n", name, h.finish().0, bytes.len()));
+        }
+        std::fs::write(ev.join("MANIFEST.tsv"), manifest).unwrap();
+    }
+
+    #[test]
+    fn honest_evidence_attests() {
+        let dir = std::env::temp_dir().join("magicka_attest_ok");
+        write_full_evidence(&dir, &[10, 20, 30, 40, 50], None);
+        let att = verify_dir(&dir).unwrap();
+        assert!(att.ok, "honest evidence should attest: {:?}", att.checks);
+        assert_eq!(att.traces_verified, att.traces_total);
+        assert!(att.traces_total > 0);
+    }
+
+    /// Negative control: a tampered leaf changes the recomputed root, so the
+    /// claimed root no longer matches and attestation FAILS.
+    #[test]
+    fn tampered_leaf_breaks_attestation() {
+        let dir = std::env::temp_dir().join("magicka_attest_bad_leaf");
+        write_full_evidence(&dir, &[10, 20, 30, 40, 50], None);
+        // Flip a leaf in the file after the root was claimed.
+        let lp = dir.join("evidence/leaves.tsv");
+        let txt = std::fs::read_to_string(&lp).unwrap();
+        let mut lines: Vec<String> = txt.lines().map(|s| s.to_string()).collect();
+        lines[2] = format!("{:016x}", 0xdead_beefu64);
+        std::fs::write(&lp, lines.join("\n")).unwrap();
+        let att = verify_dir(&dir).unwrap();
+        assert!(!att.ok, "tampered leaf must fail attestation");
+        assert!(att.checks.iter().any(|(n, ok)| n.contains("root") && !ok));
+    }
+
+    /// Negative control for finding 4: corrupting the FULL TRACE (without
+    /// touching the claimed leaf) makes the recomputed leaf disagree, so the
+    /// trace no longer attests. Proves the evidence is the full trace, not a
+    /// trusted digest.
+    #[test]
+    fn tampered_trace_breaks_attestation() {
+        let dir = std::env::temp_dir().join("magicka_attest_bad_trace");
+        write_full_evidence(&dir, &[10, 20, 30, 40, 50], Some(2));
+        let att = verify_dir(&dir).unwrap();
+        assert!(!att.ok, "tampered full trace must fail attestation");
+        assert!(att.traces_verified < att.traces_total);
+        assert!(att.checks.iter().any(|(n, ok)| n.contains("full trace") && !ok));
+    }
+}

crates/attestation/src/main.rs

@@ -0,0 +1,48 @@
+//! `attest` — independently verify a CI run's evidence directory.
+//!
+//! Usage: `attest <dir>` (default `ci_out`). Exits non-zero if the recomputed
+//! Merkle root does not match the claimed root, the leaf count is inconsistent,
+//! or the leaves are missing. It deliberately ignores `compliance_report.json`.
+
+use std::path::PathBuf;
+
+fn main() {
+    let dir = std::env::args().nth(1).unwrap_or_else(|| "ci_out".to_string());
+    let path = PathBuf::from(&dir);
+    match attestation::verify_dir(&path) {
+        Ok(att) => {
+            eprintln!("=== independent attestation of {dir} ===");
+            eprintln!("  leaves: {}", att.leaf_count);
+            eprintln!("  recomputed root: {:016x}", att.recomputed_root);
+            match att.claimed_root {
+                Some(r) => eprintln!("  claimed root:    {:016x}", r),
+                None => eprintln!("  claimed root:    <missing>"),
+            }
+            eprintln!(
+                "  full traces verified: {}/{} (reconstructed + leaf re-derived)",
+                att.traces_verified, att.traces_total
+            );
+            eprintln!(
+                "  causal records recomputed: {}/{} ({} confirmed)",
+                att.causal_recomputed, att.causal_total, att.causal_confirmed
+            );
+            eprintln!(
+                "  collapse rows derived from full traces: {}/{}",
+                att.collapse_derived, att.collapse_total
+            );
+            for (name, ok) in &att.checks {
+                eprintln!("  [{}] {}", if *ok { "PASS" } else { "FAIL" }, name);
+            }
+            if att.ok {
+                eprintln!("ATTESTATION: PASS — leaves recompute to the claimed root.");
+            } else {
+                eprintln!("ATTESTATION: FAIL — evidence is inconsistent.");
+                std::process::exit(1);
+            }
+        }
+        Err(e) => {
+            eprintln!("ATTESTATION: ERROR — {e}");
+            std::process::exit(2);
+        }
+    }
+}

crates/ci_reports/src/main.rs

@@ -1,9 +1,10 @@
 //! The `ci` binary: runs the full adversarial framework against the reference
-//! and runtime-under-test, writes the eight required reports (JSON + markdown),
+//! and the independently-implemented runtime under test, writes the required
-//! and exits nonzero if any gate fails.
+//! reports (JSON + markdown) including a provenance report binding the numbers
+//! to executed work, and exits nonzero if any gate fails.
 
 use ci_reports::json::Json;
-use ci_reports::{run_all, CiResults, Scale};
+use ci_reports::{run_all_to, CiResults, Profile, Scale};
 use std::fs;
 use std::io::Write;
 use std::path::Path;
@@ -18,6 +19,9 @@ fn fails(v: &[String]) -> Json {
 fn pass_field(v: &[String]) -> Json {
     Json::Bool(v.is_empty())
 }
+fn hex(h: world_model::Hash) -> Json {
+    Json::s(format!("{:016x}", h.0))
+}
 
 fn write_report(dir: &Path, name: &str, j: &Json) {
     let path = dir.join(format!("{name}.json"));
@@ -25,6 +29,93 @@ fn write_report(dir: &Path, name: &str, j: &Json) {
     f.write_all(j.to_pretty().as_bytes()).expect("write report");
 }
 
+/// Write the raw evidence the independent `attest` binary verifies (findings
+/// 2, 3): every Merkle leaf, plus the producer's claims. The attestor recomputes
+/// the root from these leaves and checks it against the claimed root, in a
+/// separate process that never reads the compliance report.
+fn write_evidence(dir: &Path, r: &CiResults) {
+    let ev = dir.join("evidence");
+    fs::create_dir_all(&ev).expect("create evidence dir");
+
+    let mut leaves = String::from("leaf\n");
+    for h in &r.merkle_leaves {
+        leaves.push_str(&format!("{:016x}\n", h.0));
+    }
+    fs::write(ev.join("leaves.tsv"), leaves).expect("write leaves");
+
+    let claims = format!(
+        "# evidence claims for independent attestation\n\
+         root\t{:016x}\n\
+         leaf_count\t{}\n\
+         total_comparisons\t{}\n\
+         reference_engine_id\t{:016x}\n\
+         rut_engine_id\t{:016x}\n\
+         engines_agree\t{}\n",
+        r.provenance.execution_merkle_root.0,
+        r.provenance.merkle_leaf_count,
+        r.provenance.total_comparisons,
+        r.provenance.reference_engine_id.0,
+        r.provenance.rut_engine_id.0,
+        r.provenance.engines_agree,
+    );
+    fs::write(ev.join("claims.tsv"), claims).expect("write claims");
+
+    // Per-edge causal intervention evidence (finding 7): one recomputable record
+    // per tested edge.
+    let mut causal = String::from(
+        "seed\tfrom_domain\tfrom_lane\tfrom_hidden\tto_domain\tto_lane\tto_hidden\tbase_dv\talt_dv\n",
+    );
+    for rec in &r.causal_explanation.evidence {
+        causal.push_str(&format!(
+            "{:016x}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\t{}\n",
+            rec.case_seed,
+            rec.from_domain,
+            rec.from_lane,
+            rec.from_hidden as u8,
+            rec.to_domain,
+            rec.to_lane,
+            rec.to_hidden as u8,
+            rec.base_dv,
+            rec.alt_dv,
+        ));
+    }
+    fs::write(ev.join("causal_evidence.tsv"), causal).expect("write causal evidence");
+
+    // Collapse feature rows (finding 4): the summaries the collapse gate consumes,
+    // in the same order as the first base trace records. Stored as f64 bits so an
+    // attestor can recompute each row from the full trace and compare bit-exact.
+    let mut crows = String::new();
+    for row in &r.collapse_feature_rows {
+        let cells: Vec<String> = row.iter().map(|v| v.to_bits().to_string()).collect();
+        crows.push_str(&cells.join(" "));
+        crows.push('\n');
+    }
+    fs::write(ev.join("collapse_feature_rows.tsv"), crows).expect("write collapse rows");
+
+    // Artifact bundle manifest (finding 1): the merge-scale claim is only valid
+    // if this complete, content-hashed bundle is retained. The attestor verifies
+    // every listed file exists and its hash + length match.
+    let bundle = [
+        "leaves.tsv",
+        "claims.tsv",
+        "causal_evidence.tsv",
+        "collapse_feature_rows.tsv",
+        "traces.tsv",
+    ];
+    let mut manifest = String::from("file\thash\tbytes\n");
+    for name in bundle {
+        let bytes = fs::read(ev.join(name)).unwrap_or_default();
+        let mut h = world_model::Hasher::new();
+        h.write_tag("evidence-file");
+        h.write_bytes(&bytes);
+        manifest.push_str(&format!("{}\t{:016x}\t{}\n", name, h.finish().0, bytes.len()));
+    }
+    fs::write(ev.join("MANIFEST.tsv"), manifest).expect("write manifest");
+    // Note: evidence/traces.tsv (the FULL per-execution trace corpus) is streamed
+    // during the run in main(), covering 100% of base executions — not written
+    // here from a capped sample.
+}
+
 fn build_reports(dir: &Path, r: &CiResults) {
     // 1. domain_participation_report
     write_report(
@@ -74,7 +165,7 @@ fn build_reports(dir: &Path, r: &CiResults) {
         .map(|rep| {
             Json::Obj(vec![
                 ("attack".into(), Json::s(rep.attack.clone())),
-                ("predicts".into(), Json::Num(rep.predicts)),
+                ("reconstructs".into(), Json::Num(rep.predicts)),
                 ("info_loss".into(), Json::Num(rep.info_loss)),
                 ("detail".into(), Json::s(rep.detail.clone())),
             ])
@@ -85,6 +176,7 @@ fn build_reports(dir: &Path, r: &CiResults) {
         "compression_resistance_report",
         &Json::Obj(vec![
             ("pass".into(), pass_field(&r.collapse.failures)),
+            ("measures".into(), Json::s("real serialized trace structure")),
             ("best_1factor".into(), Json::Num(r.collapse.best_1factor)),
             ("best_2factor".into(), Json::Num(r.collapse.best_2factor)),
             ("best_4factor".into(), Json::Num(r.collapse.best_4factor)),
@@ -106,20 +198,38 @@ fn build_reports(dir: &Path, r: &CiResults) {
             ("altered_trace".into(), Json::Num(r.metamorphic.altered_trace)),
             ("altered_delta".into(), Json::Num(r.metamorphic.altered_delta)),
             ("altered_future".into(), Json::Num(r.metamorphic.altered_future)),
-            ("neutral_unexplained".into(), Json::Num(r.metamorphic.neutral_unexplained)),
+            ("consumed_perturbation_trace_violations".into(), Json::Num(r.metamorphic.expectation_violations)),
+            ("consumed_delta_change_rate".into(), Json::Num(r.metamorphic.consumed_delta_rate)),
+            ("consumed_future_change_rate".into(), Json::Num(r.metamorphic.consumed_future_rate)),
+            ("legitimately_neutral".into(), Json::Num(r.metamorphic.explained_neutral)),
+            ("consumed_perturbations".into(), Json::Int(r.metamorphic.consumed as i64)),
             ("failures".into(), fails(&r.metamorphic.failures)),
         ]),
     );
 
+    // 4b. causal_explanation_report (finding 10: intervention-confirmed edges)
+    write_report(
+        dir,
+        "causal_explanation_report",
+        &Json::Obj(vec![
+            ("pass".into(), pass_field(&r.causal_explanation.failures)),
+            ("edges_tested".into(), Json::Int(r.causal_explanation.edges_tested as i64)),
+            ("edges_confirmed".into(), Json::Int(r.causal_explanation.edges_confirmed as i64)),
+            ("confirmed_fraction".into(), Json::Num(r.causal_explanation.confirmed_fraction)),
+            ("method".into(), Json::s("ablate recorded causal source lane; require destination delta to change")),
+            ("failures".into(), fails(&r.causal_explanation.failures)),
+        ]),
+    );
+
     // 5. mutation_survivor_report
     let survivors: Vec<Json> = r
         .mutation
         .survivors
         .iter()
-        .map(|(id, name)| {
+        .map(|(id, reason)| {
             Json::Obj(vec![
                 ("id".into(), Json::Int(*id as i64)),
-                ("name".into(), Json::s(name.clone())),
+                ("reason".into(), Json::s(reason.clone())),
             ])
         })
         .collect();
@@ -128,6 +238,7 @@ fn build_reports(dir: &Path, r: &CiResults) {
         "mutation_survivor_report",
         &Json::Obj(vec![
             ("pass".into(), Json::Bool(r.mutation.passed())),
+            ("killed_by".into(), Json::s("named acceptance gate")),
             ("total_mutants".into(), Json::Int(r.mutation.total as i64)),
             ("killed".into(), Json::Int(r.mutation.killed as i64)),
             ("survivors".into(), Json::Arr(survivors)),
@@ -140,6 +251,10 @@ fn build_reports(dir: &Path, r: &CiResults) {
         "runtime_equivalence_report",
         &Json::Obj(vec![
             ("pass".into(), pass_field(&r.equivalence.failures)),
+            ("independent_implementations".into(), Json::Bool(r.equivalence.independent)),
+            ("reference_engine_id".into(), hex(r.provenance.reference_engine_id)),
+            ("rut_engine_id".into(), hex(r.provenance.rut_engine_id)),
+            ("engines_agree".into(), Json::Bool(r.provenance.engines_agree)),
             ("total".into(), Json::Int(r.equivalence.total as i64)),
             ("matched".into(), Json::Int(r.equivalence.matched as i64)),
             ("failures".into(), fails(&r.equivalence.failures)),
@@ -152,9 +267,13 @@ fn build_reports(dir: &Path, r: &CiResults) {
         "replay_report",
         &Json::Obj(vec![
             ("pass".into(), pass_field(&r.replay.failures)),
+            ("loaded_from_committed_corpus".into(), Json::Bool(r.replay.loaded_from_disk)),
             ("total".into(), Json::Int(r.replay.total as i64)),
             ("deterministic".into(), Json::Int(r.replay.deterministic as i64)),
             ("drift".into(), Json::Int(r.replay.drift as i64)),
+            ("retained_failures_present".into(), Json::Bool(r.replay.retained_present)),
+            ("retained_failures_total".into(), Json::Int(r.replay.retained_total as i64)),
+            ("retained_failures_regressions".into(), Json::Int(r.replay.retained_regressions as i64)),
             ("failures".into(), fails(&r.replay.failures)),
         ]),
     );
@@ -175,6 +294,238 @@ fn build_reports(dir: &Path, r: &CiResults) {
             ("failures".into(), fails(&r.coverage.failures)),
         ]),
     );
+
+    // 9. provenance_report — binds the run to executed work.
+    let p = &r.provenance;
+    write_report(
+        dir,
+        "provenance_report",
+        &Json::Obj(vec![
+            ("pass".into(), pass_field(&p.failures)),
+            ("profile".into(), Json::s(p.profile.name())),
+            (
+                "merge_floor".into(),
+                Json::Obj(vec![
+                    ("worlds".into(), Json::Int(p.floor.worlds as i64)),
+                    ("programs".into(), Json::Int(p.floor.programs as i64)),
+                    ("executions".into(), Json::Int(p.floor.executions as i64)),
+                    ("perturbations_per_exec".into(), Json::Int(p.floor.perturbations_per_exec as i64)),
+                    ("mutants".into(), Json::Int(p.floor.mutants as i64)),
+                    ("replay_cases".into(), Json::Int(p.floor.replay_cases as i64)),
+                ]),
+            ),
+            ("worlds_generated".into(), Json::Int(p.worlds_generated as i64)),
+            ("programs_generated".into(), Json::Int(p.programs_generated as i64)),
+            ("actual_executions".into(), Json::Int(p.actual_executions as i64)),
+            ("min_perturbations_per_exec".into(), Json::Int(p.min_perturbations_per_exec as i64)),
+            ("total_comparisons".into(), Json::Int(p.total_comparisons as i64)),
+            ("actual_mutants".into(), Json::Int(p.actual_mutants as i64)),
+            ("replay_total".into(), Json::Int(p.replay_total as i64)),
+            ("reference_engine_id".into(), hex(p.reference_engine_id)),
+            ("rut_engine_id".into(), hex(p.rut_engine_id)),
+            ("engines_agree".into(), Json::Bool(p.engines_agree)),
+            ("execution_merkle_root".into(), hex(p.execution_merkle_root)),
+            ("merkle_leaf_count".into(), Json::Int(p.merkle_leaf_count as i64)),
+            ("collapse_feature_width".into(), Json::Int(p.collapse_feature_width as i64)),
+            ("failures".into(), fails(&p.failures)),
+        ]),
+    );
+}
+
+/// The reports the spec requires every CI run to produce. A missing or empty
+/// artifact is itself an acceptance failure (compliance rule 4).
+const REQUIRED_REPORTS: [&str; 10] = [
+    "domain_participation_report",
+    "causal_rank_report",
+    "causal_explanation_report",
+    "compression_resistance_report",
+    "metamorphic_response_report",
+    "mutation_survivor_report",
+    "runtime_equivalence_report",
+    "replay_report",
+    "coverage_report",
+    "provenance_report",
+];
+
+fn report_present(dir: &Path, name: &str) -> bool {
+    fs::metadata(dir.join(format!("{name}.json")))
+        .map(|m| m.len() > 0)
+        .unwrap_or(false)
+}
+
+/// One acceptance obligation: a measured artifact, a provenance chain, a
+/// merge-blocking enforcement point, and a failure if the artifact/provenance is
+/// absent. This is the compliance model made machine-checkable.
+struct Obligation {
+    requirement: &'static str,
+    artifact: &'static str,
+    floor: i64,
+    actual: i64,
+    gate_pass: bool,
+}
+
+fn obligations(r: &CiResults) -> Vec<Obligation> {
+    let p = &r.provenance;
+    let merge = p.profile == Profile::Merge;
+    let f = |v: usize| v as i64;
+    let floor = |v: usize| if merge { v as i64 } else { 0 };
+    vec![
+        Obligation {
+            requirement: "generated worlds >= 50,000",
+            artifact: "provenance_report",
+            floor: floor(p.floor.worlds),
+            actual: f(p.worlds_generated),
+            gate_pass: !merge || p.worlds_generated >= p.floor.worlds,
+        },
+        Obligation {
+            requirement: "generated programs >= 250,000",
+            artifact: "provenance_report",
+            floor: floor(p.floor.programs),
+            actual: f(p.programs_generated),
+            gate_pass: !merge || p.programs_generated >= p.floor.programs,
+        },
+        Obligation {
+            requirement: "executions >= 1,000,000",
+            artifact: "provenance_report",
+            floor: floor(p.floor.executions),
+            actual: f(p.actual_executions),
+            gate_pass: !merge || p.actual_executions >= p.floor.executions,
+        },
+        Obligation {
+            requirement: "perturbations per execution >= 10",
+            artifact: "metamorphic_response_report",
+            floor: floor(p.floor.perturbations_per_exec),
+            actual: f(p.min_perturbations_per_exec),
+            gate_pass: !merge || p.min_perturbations_per_exec >= p.floor.perturbations_per_exec,
+        },
+        Obligation {
+            requirement: "reference/runtime comparison = 100% of executions",
+            artifact: "runtime_equivalence_report",
+            floor: floor(p.floor.executions * (1 + p.floor.perturbations_per_exec)),
+            actual: f(p.total_comparisons),
+            gate_pass: r.equivalence.failures.is_empty(),
+        },
+        Obligation {
+            requirement: "semantic mutants >= 500, 0 survivors, killed by named gate",
+            artifact: "mutation_survivor_report",
+            floor: floor(p.floor.mutants),
+            actual: f(r.mutation.total),
+            gate_pass: r.mutation.passed() && (!merge || r.mutation.total >= p.floor.mutants),
+        },
+        Obligation {
+            requirement: "replay corpus >= 10,000, persisted, 0 drift",
+            artifact: "replay_report",
+            floor: floor(p.floor.replay_cases),
+            actual: f(r.replay.total),
+            gate_pass: r.replay.failures.is_empty() && r.replay.loaded_from_disk,
+        },
+        Obligation {
+            requirement: "collapse attacks fail to simplify (real trace info)",
+            artifact: "compression_resistance_report",
+            floor: 0,
+            actual: 0,
+            gate_pass: r.collapse.failures.is_empty(),
+        },
+        Obligation {
+            requirement: "trace gates (causal edges/rank/touched/fingerprints)",
+            artifact: "causal_rank_report",
+            floor: 0,
+            actual: 0,
+            gate_pass: r.trace.failures.is_empty(),
+        },
+        Obligation {
+            requirement: "domain participation (no decorative/read-only/write-only)",
+            artifact: "domain_participation_report",
+            floor: 0,
+            actual: 0,
+            gate_pass: r.domain.failures.is_empty(),
+        },
+        Obligation {
+            requirement: "metamorphic response thresholds",
+            artifact: "metamorphic_response_report",
+            floor: 0,
+            actual: 0,
+            gate_pass: r.metamorphic.failures.is_empty(),
+        },
+        Obligation {
+            requirement: "recorded causal edges are intervention-confirmed (not counted)",
+            artifact: "causal_explanation_report",
+            floor: 0,
+            actual: f(r.causal_explanation.edges_confirmed),
+            gate_pass: r.causal_explanation.failures.is_empty(),
+        },
+        Obligation {
+            requirement: "every admitted case satisfies its contract",
+            artifact: "coverage_report",
+            floor: 0,
+            actual: 0,
+            gate_pass: r.contract.failures.is_empty() && r.coverage.failures.is_empty(),
+        },
+        Obligation {
+            requirement: "provenance binds reports to executed work",
+            artifact: "provenance_report",
+            floor: 0,
+            actual: 0,
+            gate_pass: r.provenance.failures.is_empty(),
+        },
+    ]
+}
+
+/// Write the compliance report and return whether the compliance model holds:
+/// every obligation's artifact is present and every obligation passes.
+fn build_compliance_report(dir: &Path, r: &CiResults) -> bool {
+    let items = obligations(r);
+    let mut all_ok = true;
+    let mut json_items = Vec::new();
+    for ob in &items {
+        let present = report_present(dir, ob.artifact);
+        let pass = present && ob.gate_pass;
+        if !pass {
+            all_ok = false;
+        }
+        json_items.push(Json::Obj(vec![
+            ("requirement".into(), Json::s(ob.requirement)),
+            ("measured_artifact".into(), Json::s(format!("{}.json", ob.artifact))),
+            ("artifact_present".into(), Json::Bool(present)),
+            ("provenance".into(), Json::s("provenance_report.json (merkle root + engine ids)")),
+            ("merge_blocking".into(), Json::Bool(true)),
+            ("floor".into(), Json::Int(ob.floor)),
+            ("actual".into(), Json::Int(ob.actual)),
+            ("pass".into(), Json::Bool(pass)),
+        ]));
+    }
+    // Every required report must exist and be non-empty.
+    let mut missing = Vec::new();
+    for rep in REQUIRED_REPORTS {
+        if !report_present(dir, rep) {
+            missing.push(rep.to_string());
+            all_ok = false;
+        }
+    }
+    write_report(
+        dir,
+        "compliance_report",
+        &Json::Obj(vec![
+            ("pass".into(), Json::Bool(all_ok)),
+            ("profile".into(), Json::s(r.provenance.profile.name())),
+            (
+                "merge_blocking_run".into(),
+                Json::Bool(r.provenance.profile == Profile::Merge),
+            ),
+            (
+                "note".into(),
+                Json::s(if r.provenance.profile == Profile::Merge {
+                    "merge profile: floors enforced, all obligations acceptance-blocking"
+                } else {
+                    "advisory profile: NOT acceptance; floors not enforced (fast/tiny)"
+                }),
+            ),
+            ("required_reports".into(), Json::Arr(REQUIRED_REPORTS.iter().map(|s| Json::s(*s)).collect())),
+            ("missing_reports".into(), Json::Arr(missing.iter().map(|s| Json::s(s.clone())).collect())),
+            ("obligations".into(), Json::Arr(json_items)),
+        ]),
+    );
+    all_ok
 }
 
 fn status(v: bool) -> &'static str {
@@ -186,22 +537,38 @@ fn status(v: bool) -> &'static str {
 }
 
 fn write_markdown(dir: &Path, r: &CiResults) {
+    let p = &r.provenance;
     let mut s = String::new();
     s.push_str("# Magicka VM — Phase 0/1 CI Report\n\n");
-    s.push_str(&format!(
+    s.push_str(&format!("Overall: **{}**\n\n", status(r.passed())));
-        "Overall: **{}**\n\n",
+    s.push_str(&format!("Profile: **{}**", p.profile.name()));
-        status(r.passed())
+    if p.profile == Profile::Fast {
-    ));
+        s.push_str(" (representative slice — NOT merge-blocking)");
+    }
+    s.push_str("\n\n");
     s.push_str(&format!(
         "Scale: executions={}, mutants={}, replay={}, collapse_samples={}\n\n",
         r.scale.executions, r.scale.mutants, r.scale.replay_cases, r.scale.collapse_samples
     ));
+    s.push_str("## Provenance\n\n");
+    s.push_str(&format!("- Execution Merkle root: `{:016x}` over {} leaves\n", p.execution_merkle_root.0, p.merkle_leaf_count));
+    s.push_str(&format!(
+        "- Reference engine id: `{:016x}`; runtime-under-test engine id: `{:016x}`; agree: **{}**\n",
+        p.reference_engine_id.0, p.rut_engine_id.0, p.engines_agree
+    ));
+    s.push_str(&format!(
+        "- Actual executions: {} (merge floor {}), min perturbations/exec: {} (floor {})\n",
+        p.actual_executions, p.floor.executions, p.min_perturbations_per_exec, p.floor.perturbations_per_exec
+    ));
+    s.push_str(&format!("- Collapse measures real trace structure ({} features/trace)\n\n", p.collapse_feature_width));
+
     s.push_str("| Report | Status | Key metrics |\n|---|---|---|\n");
     s.push_str(&format!(
-        "| runtime_equivalence | {} | {}/{} matched |\n",
+        "| runtime_equivalence | {} | {}/{} matched, independent impls, engines agree={} |\n",
         status(r.equivalence.failures.is_empty()),
         r.equivalence.matched,
-        r.equivalence.total
+        r.equivalence.total,
+        p.engines_agree
     ));
     s.push_str(&format!(
         "| causal_rank/trace | {} | rank med={} p95={}, edges med={}, touched med={} |\n",
@@ -234,7 +601,7 @@ fn write_markdown(dir: &Path, r: &CiResults) {
         r.collapse.min_info_loss
     ));
     s.push_str(&format!(
-        "| mutation_survivor | {} | killed {}/{} |\n",
+        "| mutation_survivor | {} | killed {}/{} by named gate |\n",
         status(r.mutation.passed()),
         r.mutation.killed,
         r.mutation.total
@@ -246,7 +613,7 @@ fn write_markdown(dir: &Path, r: &CiResults) {
         r.contract.total
     ));
     s.push_str(&format!(
-        "| replay | {} | {}/{} deterministic, drift={} |\n",
+        "| replay | {} | {}/{} deterministic (committed corpus), drift={} |\n",
         status(r.replay.failures.is_empty()),
         r.replay.deterministic,
         r.replay.total,
@@ -259,6 +626,11 @@ fn write_markdown(dir: &Path, r: &CiResults) {
         r.coverage.perturbations,
         r.coverage.generated_rejected
     ));
+    s.push_str(&format!(
+        "| provenance | {} | merkle over {} leaves, floor enforced |\n",
+        status(r.provenance.failures.is_empty()),
+        r.provenance.merkle_leaf_count
+    ));
 
     s.push_str("\n## Failures\n\n");
     let mut any = false;
@@ -268,9 +640,9 @@ fn write_markdown(dir: &Path, r: &CiResults) {
             s.push_str(&format!("- **{}**: {}\n", name, msg));
         }
     }
-    for (id, name) in &r.mutation.survivors {
+    for (id, reason) in &r.mutation.survivors {
         any = true;
-        s.push_str(&format!("- **mutation_survivor**: mutant {} ({}) survived\n", id, name));
+        s.push_str(&format!("- **mutation_survivor**: mutant {} survived ({})\n", id, reason));
     }
     if !any {
         s.push_str("None. The fake universe failed to collapse. ✅\n");
@@ -287,15 +659,31 @@ fn main() {
     fs::create_dir_all(dir).expect("create out dir");
 
     eprintln!(
-        "running CI: executions={} mutants={} replay={} collapse_samples={}",
+        "running CI: profile={} executions={} mutants={} replay={} collapse_samples={}",
-        scale.executions, scale.mutants, scale.replay_cases, scale.collapse_samples
+        scale.profile.name(), scale.executions, scale.mutants, scale.replay_cases, scale.collapse_samples
     );
+    for v in &scale.override_violations {
+        eprintln!("  override rejected: {v}");
+    }
+    let merge = scale.profile == Profile::Merge;
     let start = Instant::now();
-    let results = run_all(scale);
+    // Stream full-trace evidence for 100% of base executions straight to disk.
+    let ev_dir = dir.join("evidence");
+    fs::create_dir_all(&ev_dir).expect("create evidence dir");
+    let traces_path = ev_dir.join("traces.tsv");
+    let mut traces_w = std::io::BufWriter::new(fs::File::create(&traces_path).expect("create traces"));
+    traces_w
+        .write_all(b"# full-trace evidence (100% of base executions): header<TAB>trace<TAB>delta\n")
+        .expect("write traces header");
+    let results = run_all_to(scale, Some(&mut traces_w));
+    traces_w.flush().expect("flush traces");
+    drop(traces_w);
     let elapsed = start.elapsed();
 
     build_reports(dir, &results);
+    write_evidence(dir, &results);
     write_markdown(dir, &results);
+    let compliance_ok = build_compliance_report(dir, &results);
 
     println!("\n=== Magicka VM CI ({:?}) ===", elapsed);
     for (name, f) in results.all_failures() {
@@ -313,13 +701,38 @@ fn main() {
             format!(", {} survivors", results.mutation.survivors.len())
         }
     );
+    println!("  {:<24} {}", "compliance", status(compliance_ok));
+    println!(
+        "  provenance: merkle={:016x} over {} leaves, engines_agree={}, comparisons={}",
+        results.provenance.execution_merkle_root.0,
+        results.provenance.merkle_leaf_count,
+        results.provenance.engines_agree,
+        results.provenance.total_comparisons,
+    );
     println!("reports written to {}/", out_dir);
 
-    if results.passed() {
+    let gates_pass = results.passed() && compliance_ok;
-        println!("\nOVERALL: PASS — the adversarial framework could not collapse the universe.");
+
+    if !merge {
+        // The fast/tiny profiles are advisory only — they may never stand in for
+        // the merge-blocking acceptance run (compliance: no default-profile
+        // substitution). Report status but make clear this is not acceptance.
+        println!(
+            "\nADVISORY ({} profile): {} — NOT a merge-blocking acceptance run. \
+             Acceptance requires MAGICKA_PROFILE=merge (full gates).",
+            results.provenance.profile.name(),
+            status(gates_pass)
+        );
+        // A failing advisory run still fails the PR check; a passing one is green
+        // but explicitly non-acceptance.
+        std::process::exit(if gates_pass { 0 } else { 1 });
+    }
+
+    if gates_pass {
+        println!("\nACCEPTANCE: PASS — full merge gates satisfied with persisted, independently reproducible evidence.");
         std::process::exit(0);
     } else {
-        println!("\nOVERALL: FAIL");
+        println!("\nACCEPTANCE: FAIL");
         std::process::exit(1);
     }
 }

crates/collapse_analysis/src/lib.rs

@@ -1,28 +1,80 @@
-//! `collapse_analysis` — the framework's compression attacks. Each attack
+//! `collapse_analysis` — the framework's compression attacks.
-//! tries to predict execution behavior with a *simpler* model. If any small
-//! model predicts above the configured thresholds, the universe has collapsed
-//! and CI must fail.
 //!
-//! The corpus is a set of (input, output) samples: `input` is the world ground
+//! Each attack tries to reconstruct the *actual serialized trace* from a
-//! truth (8 domains x (observed+hidden) lanes) and `output` is the behavior
+//! simpler (compressed) representation of it. If a small model reconstructs the
-//! feature vector produced by the runtime. Inputs/outputs are standardized so
+//! trace above the configured thresholds, the universe has collapsed and CI must
-//! that no single high-magnitude axis dominates the variance accounting.
+//! fail.
+//!
+//! The earlier version of this crate operated on a 22-element behavior
+//! *fingerprint* — a hash-derived proxy the avalanche engine guarantees is
+//! near-random — and defined information loss circularly as `1 - predicts`. That
+//! made every attack pass for free: the metric never touched real trace content.
+//!
+//! This version feeds the genuine trace structure (per-domain causal influence,
+//! information flow, access counts, temporal reach, and state deltas, plus the
+//! global rank/edge/divergence summaries) into the attacks, and defines
+//! information loss as the real unexplained-variance fraction of reconstructing
+//! the trace. The negative-control test builds a deliberately collapsible
+//! (single-factor) corpus and verifies the gate rejects it, demonstrating the
+//! gate discriminates a rich universe from a degenerate one.
 
 pub mod linalg;
 
 use linalg::{ols_r2, pca_scores, Mat};
-use world_model::{Hash, HIDDEN_LANES, LANES, NUM_DOMAINS};
+use trace_model::ExecutionTrace;
+use world_model::{WorldDelta, NUM_DOMAINS};
 
-/// Input columns belonging to one domain (observed + hidden lanes).
+/// The single definition of a trace feature row. The collapse corpus is built
-pub const DOMAIN_BLOCK: usize = LANES + HIDDEN_LANES;
+/// from these rows; an attestor recomputes the same row from the retained FULL
+/// trace and delta and checks equality, which is how the summary is proven to
+/// derive from the full trace. Layout per domain block (`BLOCK_W`):
+/// `[infl_out, infl_in, flow_out, flow_in, read, write, temporal, obs_delta,
+/// hid_delta]`, then globals `[causal_rank, edge_count, touched, divergence]`.
+pub fn trace_feature_row(trace: &ExecutionTrace, delta: &WorldDelta) -> Vec<f64> {
+    let infl = trace.causal_graph.influence_matrix();
+    let mut flow = [[0.0f64; NUM_DOMAINS]; NUM_DOMAINS];
+    for &(a, b, bits) in &trace.information_flow.edges {
+        flow[a as usize % NUM_DOMAINS][b as usize % NUM_DOMAINS] += bits as f64;
+    }
+    let mut temporal = [0.0f64; NUM_DOMAINS];
+    for &(_s, _off, d) in &trace.temporal_graph.edges {
+        temporal[d as usize % NUM_DOMAINS] += 1.0;
+    }
+    let mut row = Vec::with_capacity(FEATURE_W);
+    for d in 0..NUM_DOMAINS {
+        let infl_out: f64 = (0..NUM_DOMAINS).map(|j| infl[d][j]).sum();
+        let infl_in: f64 = (0..NUM_DOMAINS).map(|i| infl[i][d]).sum();
+        let flow_out: f64 = (0..NUM_DOMAINS).map(|j| flow[d][j]).sum();
+        let flow_in: f64 = (0..NUM_DOMAINS).map(|i| flow[i][d]).sum();
+        let read = trace.read_graph.access_count[d] as f64;
+        let write = trace.write_graph.access_count[d] as f64;
+        let temp = temporal[d];
+        let obs_delta: f64 = delta.domain_deltas[d].observed.iter().map(|&v| (v as f64).abs()).sum();
+        let hid_delta: f64 = delta.domain_deltas[d].hidden.iter().map(|&v| (v as f64).abs()).sum();
+        row.extend_from_slice(&[
+            infl_out, infl_in, flow_out, flow_in, read, write, temp, obs_delta, hid_delta,
+        ]);
+    }
+    row.push(trace.causal_rank() as f64);
+    row.push(trace.causal_edge_count() as f64);
+    row.push(trace.touched_domain_count() as f64);
+    row.push(trace.context_divergence());
+    row
+}
 
-/// A standardized behavior corpus.
+/// Width of one per-domain feature block in a trace feature row.
+/// `[infl_out, infl_in, flow_out, flow_in, read, write, temporal, obs_delta, hid_delta]`
+pub const BLOCK_W: usize = 9;
+/// Number of trailing global (non-domain) feature columns.
+/// `[causal_rank, edge_count, touched_count, divergence_mean]`
+pub const GLOBALS: usize = 4;
+/// Total trace feature width.
+pub const FEATURE_W: usize = NUM_DOMAINS * BLOCK_W + GLOBALS;
+
+/// A standardized corpus of real trace feature rows.
 pub struct BehaviorCorpus {
-    /// `n x (NUM_DOMAINS*DOMAIN_BLOCK)` standardized input.
+    /// `n x FEATURE_W` standardized trace features.
-    pub x: Mat,
+    pub traces: Mat,
-    /// `n x m` standardized output (behavior features).
-    pub y: Mat,
-    pub fingerprints: Vec<Hash>,
 }
 
 fn standardize(rows: &[Vec<f64>]) -> Mat {
@@ -34,7 +86,6 @@ fn standardize(rows: &[Vec<f64>]) -> Mat {
             m.set(r, c, rows[r][c]);
         }
     }
-    // center + scale to unit std per column
     for c in 0..cols {
         let mut mean = 0.0;
         for r in 0..n {
@@ -57,47 +108,65 @@ fn standardize(rows: &[Vec<f64>]) -> Mat {
 }
 
 impl BehaviorCorpus {
-    pub fn build(inputs: Vec<Vec<f64>>, outputs: Vec<Vec<f64>>, fingerprints: Vec<Hash>) -> Self {
+    pub fn build(rows: Vec<Vec<f64>>) -> Self {
         BehaviorCorpus {
-            x: standardize(&inputs),
+            traces: standardize(&rows),
-            y: standardize(&outputs),
-            fingerprints,
         }
     }
 
     pub fn n(&self) -> usize {
-        self.x.rows
+        self.traces.rows
+    }
+
+    fn cols(&self) -> usize {
+        self.traces.cols
+    }
+
+    fn block_cols(d: usize) -> Vec<usize> {
+        (d * BLOCK_W..(d + 1) * BLOCK_W).collect()
     }
 
     fn select(&self, cols: &[usize]) -> Mat {
-        let mut m = Mat::zeros(self.x.rows, cols.len());
+        let mut m = Mat::zeros(self.traces.rows, cols.len());
-        for r in 0..self.x.rows {
+        for r in 0..self.traces.rows {
             for (j, &c) in cols.iter().enumerate() {
-                m.set(r, j, self.x.at(r, c));
+                m.set(r, j, self.traces.at(r, c));
             }
         }
         m
     }
 
-    fn domain_cols(domain: usize) -> Vec<usize> {
+    fn complement(&self, cols: &[usize]) -> Vec<usize> {
-        (domain * DOMAIN_BLOCK..(domain + 1) * DOMAIN_BLOCK).collect()
+        (0..self.cols()).filter(|c| !cols.contains(c)).collect()
     }
 
-    /// R² of predicting output from the top-`k` PCA factors of the full input.
+    /// R² of reconstructing the columns `target` from the columns `source`.
-    pub fn predict_k_factor(&self, k: usize) -> f64 {
+    fn reconstruct(&self, source: &[usize], target: &[usize]) -> f64 {
-        if self.n() == 0 {
+        if source.is_empty() || target.is_empty() {
             return 0.0;
         }
-        let scores = pca_scores(&self.x, k);
+        let x = self.select(source);
-        ols_r2(&scores, &self.y)
+        let y = self.select(target);
+        ols_r2(&x, &y)
     }
 
-    /// Best R² obtainable using just a single domain's input block.
+    /// Explained-variance fraction of the whole trace from its top-`k` PCA
+    /// factors (a genuine k-factor reconstruction quality).
+    pub fn predict_k_factor(&self, k: usize) -> f64 {
+        if self.n() == 0 || self.cols() == 0 {
+            return 0.0;
+        }
+        let scores = pca_scores(&self.traces, k);
+        ols_r2(&scores, &self.traces)
+    }
+
+    /// Best fraction of the *rest of the trace* explained by a single domain.
     pub fn max_single_domain(&self) -> (usize, f64) {
         let mut best = (0usize, 0.0);
         for d in 0..NUM_DOMAINS {
-            let sub = self.select(&Self::domain_cols(d));
+            let src = Self::block_cols(d);
-            let r2 = ols_r2(&sub, &self.y);
+            let tgt = self.complement(&src);
+            let r2 = self.reconstruct(&src, &tgt);
             if r2 > best.1 {
                 best = (d, r2);
             }
@@ -105,15 +174,15 @@ impl BehaviorCorpus {
         best
     }
 
-    /// Best R² obtainable using any pair of domain blocks.
+    /// Best fraction of the rest explained by any pair of domains.
     pub fn max_pair(&self) -> ((usize, usize), f64) {
         let mut best = ((0usize, 1usize), 0.0);
         for a in 0..NUM_DOMAINS {
             for b in (a + 1)..NUM_DOMAINS {
-                let mut cols = Self::domain_cols(a);
+                let mut src = Self::block_cols(a);
-                cols.extend(Self::domain_cols(b));
+                src.extend(Self::block_cols(b));
-                let sub = self.select(&cols);
+                let tgt = self.complement(&src);
-                let r2 = ols_r2(&sub, &self.y);
+                let r2 = self.reconstruct(&src, &tgt);
                 if r2 > best.1 {
                     best = ((a, b), r2);
                 }
@@ -121,13 +190,9 @@ impl BehaviorCorpus {
         }
         best
     }
-
-    fn full_r2(&self) -> f64 {
-        ols_r2(&self.x, &self.y)
-    }
 }
 
-/// A compressed model's predictive power and information loss.
+/// A compressed model's reconstruction power and (genuine) information loss.
 #[derive(Clone, Debug)]
 pub struct CompressedModel {
     pub predicts: f64,
@@ -139,6 +204,15 @@ pub struct CompressedModel {
 pub trait CollapseAttack {
     fn name(&self) -> &'static str;
     fn compress(&self, corpus: &BehaviorCorpus) -> CompressedModel;
+    /// Whether this attack produces a dimensionality-reduced representation of
+    /// the *entire* trace (a genuine compressed model), as opposed to an
+    /// ablation probe that reconstructs one facet from the rest. Only
+    /// whole-trace compressors define the "compressed model loses ≥ 35%
+    /// information" gate; ablation probes are reported for the record and feed
+    /// their own structural gates (domain participation, etc.).
+    fn whole_trace(&self) -> bool {
+        false
+    }
 }
 
 /// One attack's outcome.
@@ -150,8 +224,21 @@ pub struct CollapseReport {
     pub detail: String,
 }
 
+/// Build a compressed-model result. `info_loss` is the genuine unexplained
+/// fraction of the reconstructed trace variance.
+fn model(predicts: f64, detail: &str) -> CompressedModel {
+    CompressedModel {
+        predicts,
+        info_loss: (1.0 - predicts).clamp(0.0, 1.0),
+        detail: detail.to_string(),
+    }
+}
+
 macro_rules! attack {
     ($name:ident, $label:expr, $body:expr) => {
+        attack!($name, $label, false, $body);
+    };
+    ($name:ident, $label:expr, $whole:expr, $body:expr) => {
         pub struct $name;
         impl CollapseAttack for $name {
             fn name(&self) -> &'static str {
@@ -161,130 +248,129 @@ macro_rules! attack {
                 let f: fn(&BehaviorCorpus) -> CompressedModel = $body;
                 f(corpus)
             }
+            fn whole_trace(&self) -> bool {
+                $whole
+            }
         }
     };
 }
 
-fn model(predicts: f64, detail: &str) -> CompressedModel {
-    CompressedModel {
-        predicts,
-        info_loss: (1.0 - predicts).clamp(0.0, 1.0),
-        detail: detail.to_string(),
-    }
-}
-
 attack!(DomainRemoval, "domain_removal", |c| {
-    // Best prediction achievable while *removing* each domain in turn.
+    // Can the rest of the trace reconstruct each removed domain's own block?
     let mut best = 0.0;
     for d in 0..NUM_DOMAINS {
-        let cols: Vec<usize> = (0..NUM_DOMAINS)
+        let tgt = BehaviorCorpus::block_cols(d);
-            .filter(|&x| x != d)
+        let src = c.complement(&tgt);
-            .flat_map(BehaviorCorpus::domain_cols)
+        best = f64::max(best, c.reconstruct(&src, &tgt));
-            .collect();
-        let sub = c.select(&cols);
-        best = f64::max(best, ols_r2(&sub, &c.y));
     }
-    model(best, "predict with one domain removed")
+    model(best, "reconstruct a removed domain from the rest")
 });
 
 attack!(DomainMerging, "domain_merging", |c| {
-    // Merge each pair into a summed block; best prediction over pairs.
+    // Merge each pair (sum their blocks); can the merged sum reconstruct the two
+    // separate blocks? If so, distinguishing the domains is redundant.
     let mut best = 0.0;
     for a in 0..NUM_DOMAINS {
         for b in (a + 1)..NUM_DOMAINS {
-            let mut merged = Mat::zeros(c.x.rows, (NUM_DOMAINS - 1) * DOMAIN_BLOCK);
+            let mut merged = Mat::zeros(c.traces.rows, BLOCK_W);
-            for r in 0..c.x.rows {
+            for r in 0..c.traces.rows {
-                let mut out_col = 0;
+                for l in 0..BLOCK_W {
-                for d in 0..NUM_DOMAINS {
+                    let v = c.traces.at(r, a * BLOCK_W + l) + c.traces.at(r, b * BLOCK_W + l);
-                    if d == b {
+                    merged.set(r, l, v);
-                        continue;
-                    }
-                    for l in 0..DOMAIN_BLOCK {
-                        let mut v = c.x.at(r, d * DOMAIN_BLOCK + l);
-                        if d == a {
-                            v += c.x.at(r, b * DOMAIN_BLOCK + l);
-                        }
-                        merged.set(r, out_col, v);
-                        out_col += 1;
-                    }
                 }
             }
-            best = f64::max(best, ols_r2(&merged, &c.y));
+            let mut tgt_cols = BehaviorCorpus::block_cols(a);
+            tgt_cols.extend(BehaviorCorpus::block_cols(b));
+            let tgt = c.select(&tgt_cols);
+            best = f64::max(best, ols_r2(&merged, &tgt));
         }
     }
-    model(best, "predict with two domains merged")
+    model(best, "reconstruct two domains from their merged sum")
 });
 
-attack!(ConstantFolding, "constant_folding", |_c| {
+attack!(ConstantFolding, "constant_folding", true, |_c| {
-    // Folding the world to constants leaves no predictive features at all.
+    // Folding the trace to constants reconstructs nothing.
-    model(0.0, "world folded to constants")
+    model(0.0, "trace folded to constants")
 });
 
 attack!(CausalEdgeDeletion, "causal_edge_deletion", |c| {
-    // Keep only each domain's first observed lane (no cross-domain structure).
+    // Drop all cross-domain causal/flow columns; can local features
-    let cols: Vec<usize> = (0..NUM_DOMAINS).map(|d| d * DOMAIN_BLOCK).collect();
+    // (counts/deltas) reconstruct the deleted causal structure?
-    let sub = c.select(&cols);
+    let mut deleted = Vec::new();
-    model(ols_r2(&sub, &c.y), "diagonal-only features")
+    for d in 0..NUM_DOMAINS {
+        for l in 0..4 {
+            // infl_out, infl_in, flow_out, flow_in
+            deleted.push(d * BLOCK_W + l);
+        }
+    }
+    let kept = c.complement(&deleted);
+    model(c.reconstruct(&kept, &deleted), "reconstruct deleted causal edges from local features")
 });
 
-attack!(StateAliasing, "state_aliasing", |c| {
+attack!(StateAliasing, "state_aliasing", true, |c| {
-    // Alias all domains into a single aggregate column.
+    // Alias the whole trace into one aggregate column; reconstruct the full trace.
-    let mut agg = Mat::zeros(c.x.rows, 1);
+    let mut agg = Mat::zeros(c.traces.rows, 1);
-    for r in 0..c.x.rows {
+    for r in 0..c.traces.rows {
         let mut s = 0.0;
-        for col in 0..c.x.cols {
+        for col in 0..c.traces.cols {
-            s += c.x.at(r, col);
+            s += c.traces.at(r, col);
         }
         agg.set(r, 0, s);
     }
-    model(ols_r2(&agg, &c.y), "single aliased aggregate")
+    model(ols_r2(&agg, &c.traces), "reconstruct trace from a single aliased aggregate")
 });
 
-attack!(LatentFactorModeling, "latent_factor_modeling", |c| {
+attack!(LatentFactorModeling, "latent_factor_modeling", true, |c| {
     model(c.predict_k_factor(4), "top-4 latent factors")
 });
 
-attack!(BehaviorClustering, "behavior_clustering", |c| {
+attack!(BehaviorClustering, "behavior_clustering", true, |c| {
-    let recon = kmeans_reconstruct(&c.y, 4);
+    let recon = kmeans_reconstruct(&c.traces, 4);
-    let scores = pca_scores(&c.x, 4);
+    model(reconstruction_r2(&c.traces, &recon), "4-cluster behavior model")
-    model(ols_r2(&scores, &recon), "4-cluster behavior model")
 });
 
-attack!(SurrogatePrediction, "surrogate_prediction", |c| {
+attack!(SurrogatePrediction, "surrogate_prediction", true, |c| {
-    model(c.full_r2(), "full linear surrogate")
+    // A small (2-factor) linear surrogate of the whole trace.
+    model(c.predict_k_factor(2), "2-factor linear surrogate")
 });
 
 attack!(TemporalFlattening, "temporal_flattening", |c| {
-    // Drop hidden lanes (time-carrying state); predict from observed only.
+    // Drop temporal columns; can the rest reconstruct temporal reach?
-    let cols: Vec<usize> = (0..NUM_DOMAINS)
+    let temporal: Vec<usize> = (0..NUM_DOMAINS).map(|d| d * BLOCK_W + 6).collect();
-        .flat_map(|d| (0..LANES).map(move |l| d * DOMAIN_BLOCK + l))
+    let kept = c.complement(&temporal);
-        .collect();
+    model(c.reconstruct(&kept, &temporal), "reconstruct temporal reach from non-temporal features")
-    let sub = c.select(&cols);
-    model(ols_r2(&sub, &c.y), "time-flattened (observed lanes only)")
 });
 
 attack!(ObservationFlattening, "observation_flattening", |c| {
-    // Use only hidden lanes (collapse the observed surface).
+    // Drop hidden-state delta columns; reconstruct them from the observed side.
-    let cols: Vec<usize> = (0..NUM_DOMAINS)
+    let hidden: Vec<usize> = (0..NUM_DOMAINS).map(|d| d * BLOCK_W + 8).collect();
-        .flat_map(|d| (0..HIDDEN_LANES).map(move |l| d * DOMAIN_BLOCK + LANES + l))
+    let kept = c.complement(&hidden);
-        .collect();
+    model(c.reconstruct(&kept, &hidden), "reconstruct hidden deltas from observed features")
-    let sub = c.select(&cols);
-    model(ols_r2(&sub, &c.y), "observation-flattened (hidden lanes only)")
 });
 
 attack!(ExecutorIdentityErasure, "executor_identity_erasure", |c| {
-    // Predict all-but-last output dim (the divergence summary) from input.
+    // The divergence summary is the last global column; reconstruct it from the
-    if c.y.cols <= 1 {
+    // rest (erasing executor identity).
-        return model(0.0, "no executor dim");
+    let div = vec![c.cols() - 1];
-    }
+    let kept = c.complement(&div);
-    let mut y2 = Mat::zeros(c.y.rows, c.y.cols - 1);
+    model(c.reconstruct(&kept, &div), "reconstruct executor divergence from the rest")
-    for r in 0..c.y.rows {
+});
-        for col in 0..c.y.cols - 1 {
+
-            y2.set(r, col, c.y.at(r, col));
+fn reconstruction_r2(original: &Mat, recon: &Mat) -> f64 {
+    let mut ss_res = 0.0;
+    let mut ss_tot = 0.0;
+    for r in 0..original.rows {
+        for col in 0..original.cols {
+            let o = original.at(r, col);
+            ss_res += (o - recon.at(r, col)).powi(2);
+            ss_tot += o.powi(2);
         }
     }
-    model(ols_r2(&c.x, &y2), "executor identity erased")
+    if ss_tot <= 1e-9 {
-});
+        return 0.0;
+    }
+    (1.0 - ss_res / ss_tot).clamp(0.0, 1.0)
+}
 
 fn kmeans_reconstruct(y: &Mat, k: usize) -> Mat {
     let n = y.rows;
@@ -292,11 +378,11 @@ fn kmeans_reconstruct(y: &Mat, k: usize) -> Mat {
         return y.clone();
     }
     let k = k.min(n);
-    // deterministic init: spread initial centroids across the data
+    let mut centroids: Vec<Vec<f64>> = (0..k)
-    let mut centroids: Vec<Vec<f64>> = (0..k).map(|i| y.data[(i * n / k) * y.cols..(i * n / k) * y.cols + y.cols].to_vec()).collect();
+        .map(|i| y.data[(i * n / k) * y.cols..(i * n / k) * y.cols + y.cols].to_vec())
+        .collect();
     let mut assign = vec![0usize; n];
     for _ in 0..12 {
-        // assign
         for r in 0..n {
             let mut best = 0;
             let mut bestd = f64::MAX;
@@ -312,7 +398,6 @@ fn kmeans_reconstruct(y: &Mat, k: usize) -> Mat {
             }
             assign[r] = best;
         }
-        // update
         let mut sums = vec![vec![0.0; y.cols]; k];
         let mut counts = vec![0usize; k];
         for r in 0..n {
@@ -377,10 +462,16 @@ impl CollapseSummary {
 /// Run every attack and check all collapse gates.
 pub fn analyze(corpus: &BehaviorCorpus) -> CollapseSummary {
     let mut reports = Vec::new();
+    // `min_info_loss` is defined only over whole-trace compressors — the genuine
+    // "compressed model loses ≥ 35% information" gate. Ablation probes run and
+    // report, but reconstructing one facet from the rest is redundancy, not
+    // whole-trace compression, and is gated structurally elsewhere.
     let mut min_info_loss: f64 = 1.0;
     for atk in all_attacks() {
         let m = atk.compress(corpus);
-        min_info_loss = min_info_loss.min(m.info_loss);
+        if atk.whole_trace() {
+            min_info_loss = min_info_loss.min(m.info_loss);
+        }
         reports.push(CollapseReport {
             attack: atk.name().to_string(),
             predicts: m.predicts,
@@ -397,13 +488,13 @@ pub fn analyze(corpus: &BehaviorCorpus) -> CollapseSummary {
 
     let mut failures = Vec::new();
     if best_1 >= 0.40 {
-        failures.push(format!("1-factor predicts {:.3} >= 0.40", best_1));
+        failures.push(format!("1-factor reconstructs {:.3} >= 0.40", best_1));
     }
     if best_2 >= 0.55 {
-        failures.push(format!("2-factor predicts {:.3} >= 0.55", best_2));
+        failures.push(format!("2-factor reconstructs {:.3} >= 0.55", best_2));
     }
     if best_4 >= 0.70 {
-        failures.push(format!("4-factor predicts {:.3} >= 0.70", best_4));
+        failures.push(format!("4-factor reconstructs {:.3} >= 0.70", best_4));
     }
     if max_single > 0.30 {
         failures.push(format!("single domain explains {:.3} > 0.30", max_single));
@@ -430,41 +521,50 @@ pub fn analyze(corpus: &BehaviorCorpus) -> CollapseSummary {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use world_model::{Hasher, Rng};
+    use world_model::Rng;
+
+    fn random_trace_rows(n: usize, seed: u64) -> Vec<Vec<f64>> {
+        let mut rng = Rng::new(seed);
+        (0..n)
+            .map(|_| {
+                (0..FEATURE_W)
+                    .map(|_| rng.range_i64(-5000, 5000) as f64)
+                    .collect()
+            })
+            .collect()
+    }
 
     #[test]
-    fn random_behavior_resists_collapse() {
+    fn high_entropy_trace_resists_collapse() {
-        // Inputs random; outputs an avalanche hash of inputs -> no small linear
+        let corpus = BehaviorCorpus::build(random_trace_rows(400, 1));
-        // model should predict them.
-        let mut rng = Rng::new(1);
-        let mut inputs = Vec::new();
-        let mut outputs = Vec::new();
-        let mut fps = Vec::new();
-        for _ in 0..400 {
-            let inp: Vec<f64> = (0..NUM_DOMAINS * DOMAIN_BLOCK)
-                .map(|_| rng.range_i64(-5000, 5000) as f64)
-                .collect();
-            let mut h = Hasher::new();
-            for &v in &inp {
-                h.write_i64(v as i64);
-            }
-            let base = h.finish().0;
-            let out: Vec<f64> = (0..12)
-                .map(|k| {
-                    let mut hh = Hasher::new();
-                    hh.write_u64(base);
-                    hh.write_u64(k);
-                    (hh.finish().0 as i64) as f64
-                })
-                .collect();
-            fps.push(world_model::Hash(base));
-            inputs.push(inp);
-            outputs.push(out);
-        }
-        let corpus = BehaviorCorpus::build(inputs, outputs, fps);
         let summary = analyze(&corpus);
         assert!(summary.passed(), "collapse failures: {:?}", summary.failures);
-        assert!(summary.best_1factor < 0.40);
+    }
-        assert!(summary.min_info_loss >= 0.35);
+
+    /// Negative control: a single-factor (rank-1) corpus is genuinely
+    /// collapsible. Every feature is a fixed loading times one latent value plus
+    /// tiny noise, so a 1-factor model reconstructs almost everything. The gate
+    /// MUST reject it — proving the collapse analysis discriminates.
+    #[test]
+    fn single_factor_corpus_is_rejected() {
+        let mut rng = Rng::new(7);
+        let loadings: Vec<f64> = (0..FEATURE_W).map(|i| 1.0 + (i % 5) as f64).collect();
+        let rows: Vec<Vec<f64>> = (0..400)
+            .map(|_| {
+                let latent = rng.range_i64(-1000, 1000) as f64;
+                loadings
+                    .iter()
+                    .map(|&load| load * latent + rng.range_i64(-2, 2) as f64)
+                    .collect()
+            })
+            .collect();
+        let corpus = BehaviorCorpus::build(rows);
+        let summary = analyze(&corpus);
+        assert!(
+            !summary.passed(),
+            "collapse gate failed to reject a single-factor universe (1f={:.3}, info_loss={:.3})",
+            summary.best_1factor,
+            summary.min_info_loss
+        );
     }
 }

crates/game_runtime/Cargo.toml

@@ -0,0 +1,17 @@
+[package]
+name = "game_runtime"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+world_model = { path = "../world_model" }
+rune_ir = { path = "../rune_ir" }
+trace_model = { path = "../trace_model" }
+# Used only for the shared engine I/O *contract* types (ResolutionInput /
+# ResolutionResult) and the canonical view. Execution is driven through the
+# independent interpreter in `runtime_under_test`, never the reference engine.
+reference_runtime = { path = "../reference_runtime" }
+runtime_under_test = { path = "../runtime_under_test" }
+generators = { path = "../generators" }
+protocol = { path = "../protocol" }

crates/game_runtime/src/lib.rs

@@ -0,0 +1,757 @@
+//! `game_runtime` — the authoritative match layer (Phase B of `plan2.md`).
+//!
+//! This crate is the *only* place game truth is decided. The browser sends
+//! intent; this crate resolves it. Every rune program executes through the
+//! **independent** interpreter [`runtime_under_test::native_resolve`] against
+//! the shared [`WorldSnapshot`] — the web layer is a window into the Rust
+//! universe and never a second simulation. The game does not call the reference
+//! engine; correctness of the interpreter it does use is established separately
+//! by the runtime-equivalence gate (which compares that interpreter against the
+//! reference over a large sweep, with a negative control proving the gate can
+//! fail).
+//!
+//! Two properties are essential and tested:
+//! * **Determinism** — a match is a pure function of `(seed, roster, ordered
+//!   inputs)`. [`replay`] reconstructs any match and produces an identical
+//!   final hash. No wall clock, no ambient RNG; the turn *timer* lives in the
+//!   server, never here.
+//! * **Authority + visibility** — players receive a [`VisibleWorldSnapshot`]
+//!   that redacts all hidden lanes and every non-observable observed lane. The
+//!   hidden ground truth is never placed in any client-bound structure.
+
+use protocol::{
+    Action, Knowledge, MatchId, RuneDiagnostics, RuneTokenWire, VisibleDomain,
+    VisibleEntity, VisibleWorldSnapshot,
+};
+use reference_runtime::{canonical, ResolutionInput, ResolutionResult};
+use runtime_under_test::native_resolve;
+use rune_ir::{Op, RuneProgram, RuneToken};
+use world_model::{
+    standard_executors, DomainKind, ExecutionContext, Hash, Hasher, ProgramId, Rng, WorldSnapshot,
+    HIDDEN_LANES, LANES, NUM_DOMAINS,
+};
+
+pub const ARENA_W: i32 = 8;
+pub const ARENA_H: i32 = 8;
+pub const MAX_HP: i32 = 30;
+/// Basic stick attack damage.
+pub const ATTACK_DAMAGE: i32 = 4;
+/// Range (Manhattan) within which a cast's consequence reaches enemies.
+pub const CAST_RANGE: i32 = 3;
+
+/// One combatant on the arena. A "player" entity is driven by a connection; a
+/// "dummy" is a deterministic stationary target for the 1-player slice.
+#[derive(Clone, Debug)]
+pub struct Entity {
+    pub id: u32,
+    pub name: String,
+    pub x: i32,
+    pub y: i32,
+    pub hp: i32,
+    pub is_dummy: bool,
+    /// The player's current editable rune program (Phase D editor state).
+    pub program: RuneProgram,
+}
+
+impl Entity {
+    pub fn alive(&self) -> bool {
+        self.hp > 0
+    }
+}
+
+/// A roster entry needed to reconstruct a match for replay.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct RosterEntry {
+    pub id: u32,
+    pub name: String,
+    pub is_dummy: bool,
+}
+
+/// One resolved turn's authoritative input, sufficient to replay it exactly.
+#[derive(Clone, Debug, PartialEq)]
+pub struct TurnInput {
+    pub player: u32,
+    pub action: Action,
+    /// The exact program used, captured iff `action` is `Cast`. Recording the
+    /// program here (rather than replaying editor edits) makes replay a pure
+    /// function of this input stream.
+    pub program: Option<Vec<RuneTokenWire>>,
+}
+
+/// One recorded turn (Phase G).
+#[derive(Clone, Debug)]
+pub struct RecordedTurn {
+    pub turn: u64,
+    pub inputs: Vec<TurnInput>,
+    pub turn_hash: Hash,
+    pub events: Vec<String>,
+}
+
+/// The full replay log for a match.
+#[derive(Clone, Debug)]
+pub struct ReplayLog {
+    pub seed: u64,
+    pub roster: Vec<RosterEntry>,
+    pub turns: Vec<RecordedTurn>,
+    pub final_hash: Hash,
+}
+
+/// The authoritative match state.
+#[derive(Clone, Debug)]
+pub struct Match {
+    pub id: MatchId,
+    pub seed: u64,
+    pub turn: u64,
+    pub world: WorldSnapshot,
+    pub contexts: Vec<ExecutionContext>,
+    pub entities: Vec<Entity>,
+    pub history: Vec<String>,
+    pub replay: ReplayLog,
+    pub finished: bool,
+    /// The most recent per-domain observed change, used to tag freshly-observed
+    /// lanes without storing per-player memory (keeps resolution stateless).
+    last_observed_delta: [[i64; LANES]; NUM_DOMAINS],
+}
+
+/// A blank default program (a single benign token) so every entity always has
+/// something to cast.
+fn default_program(seed: u64) -> RuneProgram {
+    let mut rng = Rng::derive(seed, "default-program");
+    let tokens = (0..8)
+        .map(|i| RuneToken {
+            op: Op::from_u8((i as u8).wrapping_add(rng.next_u64() as u8)),
+            a: rng.next_u64() as u8,
+            b: rng.next_u64() as u8,
+            c: rng.next_u64() as u8,
+            imm: rng.range_i64(-1000, 1000),
+        })
+        .collect();
+    RuneProgram { id: ProgramId(seed), tokens, seed }
+}
+
+impl Match {
+    /// Create a match deterministically from a seed and a roster.
+    pub fn new(id: MatchId, seed: u64, roster: Vec<RosterEntry>) -> Match {
+        let world = generators::generate_world(seed);
+        let contexts = standard_executors(seed, 3);
+        let mut placer = Rng::derive(seed, "arena-placement");
+        let mut taken: Vec<(i32, i32)> = Vec::new();
+        let mut entities = Vec::with_capacity(roster.len());
+        for entry in &roster {
+            // Deterministic distinct placement.
+            let (x, y) = loop {
+                let x = placer.below(ARENA_W as usize) as i32;
+                let y = placer.below(ARENA_H as usize) as i32;
+                if !taken.contains(&(x, y)) {
+                    break (x, y);
+                }
+            };
+            taken.push((x, y));
+            entities.push(Entity {
+                id: entry.id,
+                name: entry.name.clone(),
+                x,
+                y,
+                hp: MAX_HP,
+                is_dummy: entry.is_dummy,
+                program: default_program(seed ^ (entry.id as u64).wrapping_mul(0x9e3779b97f4a7c15)),
+            });
+        }
+        Match {
+            id,
+            seed,
+            turn: 0,
+            world,
+            contexts,
+            entities,
+            history: Vec::new(),
+            replay: ReplayLog { seed, roster, turns: Vec::new(), final_hash: Hash(0) },
+            finished: false,
+            last_observed_delta: [[0; LANES]; NUM_DOMAINS],
+        }
+    }
+
+    pub fn entity(&self, id: u32) -> Option<&Entity> {
+        self.entities.iter().find(|e| e.id == id)
+    }
+    pub fn entity_mut(&mut self, id: u32) -> Option<&mut Entity> {
+        self.entities.iter_mut().find(|e| e.id == id)
+    }
+
+    /// Replace a player's editable program (Phase D / `EditRuneProgram`). Length
+    /// is bounded by the protocol decoder; this just stores it.
+    pub fn set_program(&mut self, player: u32, tokens: Vec<RuneTokenWire>) {
+        let seed = self.seed;
+        if let Some(e) = self.entity_mut(player) {
+            e.program = RuneProgram {
+                id: ProgramId(player as u64),
+                seed: seed ^ player as u64,
+                tokens: tokens.iter().map(|t| t.into_token()).collect(),
+            };
+        }
+    }
+
+    fn resolution_input(&self, program: &RuneProgram) -> ResolutionInput {
+        ResolutionInput {
+            world: self.world.clone(),
+            program: program.clone(),
+            contexts: self.contexts.clone(),
+            contract_seed: self.seed,
+            perturbation_seed: self.seed ^ self.turn,
+        }
+    }
+
+    /// Resolve one turn from a set of `(player, action)` submissions. Missing
+    /// players default to `Wait`. Returns the per-turn event list. This is the
+    /// authoritative state transition and is fully deterministic.
+    pub fn resolve_turn(&mut self, submissions: &[(u32, Action)]) -> Vec<String> {
+        // Build a canonical, complete, sorted input set: one action per entity.
+        let mut inputs: Vec<TurnInput> = Vec::new();
+        let mut ids: Vec<u32> = self.entities.iter().map(|e| e.id).collect();
+        ids.sort_unstable();
+        for id in ids {
+            let action = submissions
+                .iter()
+                .find(|(pid, _)| *pid == id)
+                .map(|(_, a)| a.clone())
+                .unwrap_or(Action::Wait);
+            let program = if matches!(action, Action::Cast) {
+                self.entity(id)
+                    .map(|e| e.program.tokens.iter().map(RuneTokenWire::from_token).collect())
+            } else {
+                None
+            };
+            inputs.push(TurnInput { player: id, action, program });
+        }
+
+        let before = self.world.clone();
+        let mut events = Vec::new();
+
+        for input in &inputs {
+            self.apply_action(input, &mut events);
+        }
+
+        // Knowledge bookkeeping: which observed lanes changed this turn.
+        for d in 0..NUM_DOMAINS {
+            for l in 0..LANES {
+                self.last_observed_delta[d][l] =
+                    self.world.domains[d].observed[l].wrapping_sub(before.domains[d].observed[l]);
+            }
+        }
+
+        self.turn = self.turn.wrapping_add(1);
+
+        // Per-turn hash binds every effect: world state + entity state + inputs.
+        let turn_hash = self.turn_hash(&inputs);
+        for e in &events {
+            self.history.push(format!("turn {}: {}", self.turn, e));
+        }
+
+        // End condition: in a multi-player match, finish when at most one
+        // non-dummy combatant is still standing.
+        let players = self.entities.iter().filter(|e| !e.is_dummy).count();
+        let living_players = self.entities.iter().filter(|e| !e.is_dummy && e.alive()).count();
+        if players >= 2 && living_players <= 1 {
+            self.finished = true;
+        }
+
+        self.replay.turns.push(RecordedTurn {
+            turn: self.turn,
+            inputs,
+            turn_hash,
+            events: events.clone(),
+        });
+        self.recompute_final_hash();
+        events
+    }
+
+    fn apply_action(&mut self, input: &TurnInput, events: &mut Vec<String>) {
+        // Skip dead entities entirely.
+        let alive = self.entity(input.player).map(|e| e.alive()).unwrap_or(false);
+        if !alive {
+            return;
+        }
+        match &input.action {
+            Action::Wait => {}
+            Action::Move { dx, dy } => {
+                let (nx, ny) = {
+                    let e = self.entity(input.player).unwrap();
+                    (
+                        (e.x + dx.clamp(&-1, &1)).clamp(0, ARENA_W - 1),
+                        (e.y + dy.clamp(&-1, &1)).clamp(0, ARENA_H - 1),
+                    )
+                };
+                let occupied = self
+                    .entities
+                    .iter()
+                    .any(|o| o.id != input.player && o.alive() && o.x == nx && o.y == ny);
+                if !occupied {
+                    let name = self.entity(input.player).unwrap().name.clone();
+                    let e = self.entity_mut(input.player).unwrap();
+                    e.x = nx;
+                    e.y = ny;
+                    events.push(format!("{name} moved to ({nx},{ny})"));
+                }
+            }
+            Action::Attack { target } => {
+                let attacker = self.entity(input.player).unwrap().clone();
+                if let Some(t) = self.entity(*target) {
+                    let adjacent = (t.x - attacker.x).abs() <= 1 && (t.y - attacker.y).abs() <= 1;
+                    if adjacent && t.alive() && *target != input.player {
+                        let tname = t.name.clone();
+                        let te = self.entity_mut(*target).unwrap();
+                        te.hp = (te.hp - ATTACK_DAMAGE).max(0);
+                        let hp = te.hp;
+                        events.push(format!(
+                            "{} struck {} for {ATTACK_DAMAGE} ({} hp left)",
+                            attacker.name, tname, hp
+                        ));
+                        if hp == 0 {
+                            events.push(format!("{tname} fell"));
+                        }
+                    }
+                }
+            }
+            Action::Inspect { target } => {
+                if let Some(t) = self.entity(*target) {
+                    events.push(format!(
+                        "{} inspected {}",
+                        self.entity(input.player).unwrap().name,
+                        t.name
+                    ));
+                }
+            }
+            Action::Cast => {
+                let program = match &input.program {
+                    Some(toks) => RuneProgram {
+                        id: ProgramId(input.player as u64),
+                        seed: self.seed ^ input.player as u64,
+                        tokens: toks.iter().map(|t| t.into_token()).collect(),
+                    },
+                    None => self.entity(input.player).unwrap().program.clone(),
+                };
+                let res = native_resolve(&self.resolution_input(&program));
+                self.apply_resolution(&res);
+                let power = cast_power(&res);
+                let caster = self.entity(input.player).unwrap().clone();
+                events.push(format!("{} cast a rune program (power {power})", caster.name));
+                // Consequence: enemies within range take `power` damage.
+                let targets: Vec<u32> = self
+                    .entities
+                    .iter()
+                    .filter(|o| {
+                        o.id != input.player
+                            && o.alive()
+                            && (o.x - caster.x).abs() + (o.y - caster.y).abs() <= CAST_RANGE
+                    })
+                    .map(|o| o.id)
+                    .collect();
+                for tid in targets {
+                    let tname = self.entity(tid).unwrap().name.clone();
+                    let te = self.entity_mut(tid).unwrap();
+                    te.hp = (te.hp - power).max(0);
+                    let hp = te.hp;
+                    events.push(format!("{tname} took {power} from the working ({hp} hp left)"));
+                    if hp == 0 {
+                        events.push(format!("{tname} fell"));
+                    }
+                }
+            }
+        }
+    }
+
+    /// Apply a resolution's world delta to the shared world (observed + hidden).
+    fn apply_resolution(&mut self, res: &ResolutionResult) {
+        for dd in &res.delta.domain_deltas {
+            let d = dd.domain.0 as usize;
+            if d >= NUM_DOMAINS {
+                continue;
+            }
+            for l in 0..LANES {
+                self.world.domains[d].observed[l] =
+                    self.world.domains[d].observed[l].wrapping_add(dd.observed[l]);
+            }
+            for l in 0..HIDDEN_LANES {
+                self.world.domains[d].hidden[l] =
+                    self.world.domains[d].hidden[l].wrapping_add(dd.hidden[l]);
+            }
+        }
+    }
+
+    fn turn_hash(&self, inputs: &[TurnInput]) -> Hash {
+        let mut h = Hasher::new();
+        h.write_tag("game-turn");
+        h.write_u64(self.turn);
+        h.write_u64(self.world.content_hash().0);
+        for e in &self.entities {
+            h.write_u64(e.id as u64);
+            h.write_i64(e.x as i64);
+            h.write_i64(e.y as i64);
+            h.write_i64(e.hp as i64);
+        }
+        for input in inputs {
+            h.write_u64(input.player as u64);
+            hash_action(&mut h, &input.action);
+            if let Some(prog) = &input.program {
+                h.write_usize(prog.len());
+                for t in prog {
+                    h.write_u8(t.op);
+                    h.write_u8(t.a);
+                    h.write_u8(t.b);
+                    h.write_u8(t.c);
+                    h.write_i64(t.imm);
+                }
+            }
+        }
+        h.finish()
+    }
+
+    fn recompute_final_hash(&mut self) {
+        let mut h = Hasher::new();
+        h.write_tag("game-replay-final");
+        h.write_u64(self.seed);
+        for entry in &self.replay.roster {
+            h.write_u64(entry.id as u64);
+            h.write_bytes(entry.name.as_bytes());
+            h.write_u8(entry.is_dummy as u8);
+        }
+        for t in &self.replay.turns {
+            h.write_u64(t.turn_hash.0);
+        }
+        self.replay.final_hash = h.finish();
+    }
+
+    /// Hex string of the most recent turn's hash (Phase G `runtime_hash`).
+    pub fn last_turn_hash_hex(&self) -> String {
+        format!("{}", self.replay.turns.last().map(|t| t.turn_hash).unwrap_or(Hash(0)))
+    }
+
+    pub fn final_hash_hex(&self) -> String {
+        format!("{}", self.replay.final_hash)
+    }
+
+    // -- Visibility / knowledge (Phase F) -----------------------------------
+
+    /// Build the filtered snapshot for one player. Hidden lanes and non-visible
+    /// observed lanes are redacted; only their *count* is reported.
+    pub fn visible_for(&self, player: u32) -> VisibleWorldSnapshot {
+        let projection = self.world.observed_projection();
+        let mut observed_domains = Vec::with_capacity(NUM_DOMAINS);
+        let mut redactions: u32 = 0;
+        for d in 0..NUM_DOMAINS {
+            let mut observed = Vec::with_capacity(LANES);
+            let mut knowledge = Vec::with_capacity(LANES);
+            for l in 0..LANES {
+                if self.world.observation_state.visible[d][l] {
+                    observed.push(Some(projection[d * LANES + l]));
+                    knowledge.push(if self.last_observed_delta[d][l] != 0 {
+                        Knowledge::NewlyObserved
+                    } else {
+                        Knowledge::Known
+                    });
+                } else {
+                    observed.push(None);
+                    knowledge.push(Knowledge::Unknown);
+                    redactions += 1;
+                }
+            }
+            observed_domains.push(VisibleDomain {
+                index: d as u8,
+                name: DomainKind::from_index(d).name().to_string(),
+                observed,
+                knowledge,
+            });
+        }
+        // All hidden lanes are always withheld.
+        redactions += (NUM_DOMAINS * HIDDEN_LANES) as u32;
+
+        let observed_entities = self
+            .entities
+            .iter()
+            .map(|e| VisibleEntity {
+                id: e.id,
+                name: e.name.clone(),
+                x: e.x,
+                y: e.y,
+                hp: e.hp,
+                is_self: e.id == player,
+                alive: e.alive(),
+            })
+            .collect();
+
+        // Inference from *observed* volatility only — never from hidden state.
+        // Use a presence test (any lane changed) rather than summing magnitudes,
+        // which avoids overflow on wrapping deltas near i64::MIN.
+        let mut inferred_markers = Vec::new();
+        for d in 0..NUM_DOMAINS {
+            let shifted = (0..LANES).any(|l| self.last_observed_delta[d][l] != 0);
+            if shifted && self.world.observation_state.visible[d].iter().any(|&v| v) {
+                inferred_markers.push(format!(
+                    "{} shifted recently — likely volatile",
+                    DomainKind::from_index(d).name()
+                ));
+            }
+        }
+
+        let known_history: Vec<String> = self.history.iter().rev().take(8).rev().cloned().collect();
+
+        VisibleWorldSnapshot {
+            turn: self.turn,
+            arena_w: ARENA_W,
+            arena_h: ARENA_H,
+            observed_domains,
+            observed_entities,
+            observed_environment: vec![
+                format!("arena {ARENA_W}x{ARENA_H}"),
+                format!("turn {}", self.turn),
+            ],
+            known_history,
+            inferred_markers,
+            hidden_state_redactions: redactions,
+        }
+    }
+
+    /// Player-facing diagnostics for a candidate program (Phase D). A *dry run*
+    /// against a clone of the world — it mutates nothing and never reports
+    /// hidden values, only domain names, counts, and observed fault risks.
+    pub fn diagnostics_for(&self, program: &RuneProgram) -> RuneDiagnostics {
+        let res = native_resolve(&self.resolution_input(program));
+        let visible_domain = |d: usize| self.world.observation_state.visible[d].iter().any(|&v| v);
+
+        let mut known_reads = Vec::new();
+        for d in res.trace.read_graph.touched() {
+            if visible_domain(d) {
+                known_reads.push(DomainKind::from_index(d).name().to_string());
+            }
+        }
+        known_reads.sort();
+        known_reads.dedup();
+
+        let mut known_writes = Vec::new();
+        let mut unknown_listeners = 0u32;
+        for d in res.trace.write_graph.touched() {
+            if visible_domain(d) {
+                known_writes.push(DomainKind::from_index(d).name().to_string());
+            } else {
+                unknown_listeners += 1;
+            }
+        }
+        known_writes.sort();
+        known_writes.dedup();
+
+        let mut observed_risks = Vec::new();
+        let mut seen = std::collections::BTreeSet::new();
+        for f in &res.faults.faults {
+            if seen.insert(f.code.name()) {
+                observed_risks.push(format!("possible {}", f.code.name()));
+            }
+        }
+
+        let matching: Vec<String> = self
+            .history
+            .iter()
+            .filter(|h| h.contains("cast") || h.contains("working"))
+            .cloned()
+            .collect();
+        let start = matching.len().saturating_sub(4);
+        let previous_outcomes: Vec<String> = matching[start..].to_vec();
+
+        RuneDiagnostics {
+            known_reads,
+            known_writes,
+            observed_risks,
+            unknown_listeners,
+            previous_outcomes,
+        }
+    }
+
+    /// Diagnostics for a player's currently-stored program.
+    pub fn diagnostics_for_player(&self, player: u32) -> RuneDiagnostics {
+        match self.entity(player) {
+            Some(e) => self.diagnostics_for(&e.program),
+            None => RuneDiagnostics::default(),
+        }
+    }
+}
+
+/// Damage power derived from the runtime trace — the rune program's effect on
+/// the game is a function of the structure the Rust engine actually produced.
+fn cast_power(res: &ResolutionResult) -> i32 {
+    let rank = res.trace.causal_rank() as i32;
+    let touched = res.trace.touched_domain_count() as i32;
+    (1 + rank + touched / 2).clamp(1, 10)
+}
+
+fn hash_action(h: &mut Hasher, a: &Action) {
+    match a {
+        Action::Wait => h.write_u8(0),
+        Action::Move { dx, dy } => {
+            h.write_u8(1);
+            h.write_i64(*dx as i64);
+            h.write_i64(*dy as i64);
+        }
+        Action::Inspect { target } => {
+            h.write_u8(2);
+            h.write_u64(*target as u64);
+        }
+        Action::Cast => h.write_u8(3),
+        Action::Attack { target } => {
+            h.write_u8(4);
+            h.write_u64(*target as u64);
+        }
+    }
+}
+
+/// Build a standard 1-player + dummy roster.
+pub fn solo_roster(player_name: &str) -> Vec<RosterEntry> {
+    vec![
+        RosterEntry { id: 1, name: player_name.to_string(), is_dummy: false },
+        RosterEntry { id: 2, name: "training dummy".to_string(), is_dummy: true },
+    ]
+}
+
+/// Build a 2-player roster.
+pub fn duel_roster(a: &str, b: &str) -> Vec<RosterEntry> {
+    vec![
+        RosterEntry { id: 1, name: a.to_string(), is_dummy: false },
+        RosterEntry { id: 2, name: b.to_string(), is_dummy: false },
+    ]
+}
+
+/// Re-run a match from its seed, roster, and the exact recorded inputs, and
+/// return the reconstructed log. Determinism gate: this must reproduce the
+/// original `final_hash` bit-for-bit.
+pub fn replay(seed: u64, roster: &[RosterEntry], recorded: &[RecordedTurn]) -> ReplayLog {
+    let mut m = Match::new(MatchId(0), seed, roster.to_vec());
+    for rt in recorded {
+        // Restore each casting player's program from the record, then apply the
+        // same actions in the same order.
+        for input in &rt.inputs {
+            if let (Action::Cast, Some(prog)) = (&input.action, &input.program) {
+                m.set_program(input.player, prog.clone());
+            }
+        }
+        let subs: Vec<(u32, Action)> =
+            rt.inputs.iter().map(|i| (i.player, i.action.clone())).collect();
+        m.resolve_turn(&subs);
+    }
+    m.replay
+}
+
+/// Convenience: run a scripted match end-to-end and return its log. Used by the
+/// determinism tests and the headless E2E harness.
+pub fn run_scripted(
+    seed: u64,
+    roster: &[RosterEntry],
+    scripts: &[Vec<(u32, Action)>],
+) -> (Match, ReplayLog) {
+    let mut m = Match::new(MatchId(seed), seed, roster.to_vec());
+    for turn_subs in scripts {
+        m.resolve_turn(turn_subs);
+    }
+    let log = m.replay.clone();
+    (m, log)
+}
+
+/// Canonical fingerprint of a single resolution (used by integration tests to
+/// assert the game layer truly drove the independent interpreter).
+pub fn resolution_fingerprint(world: &WorldSnapshot, program: &RuneProgram, seed: u64) -> Hash {
+    let input = ResolutionInput {
+        world: world.clone(),
+        program: program.clone(),
+        contexts: standard_executors(seed, 3),
+        contract_seed: seed,
+        perturbation_seed: seed,
+    };
+    let c = canonical(&native_resolve(&input));
+    let mut h = Hasher::new();
+    h.write_tag("resolution-fp");
+    h.write_u64(c.delta_hash.0);
+    h.write_u64(c.trace_hash.0);
+    h.write_u64(c.replay_hash.0);
+    h.finish()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn rune(op: u8, a: u8, b: u8, c: u8, imm: i64) -> RuneTokenWire {
+        RuneTokenWire { op, a, b, c, imm }
+    }
+
+    fn scripted_match() -> Vec<Vec<(u32, Action)>> {
+        vec![
+            vec![(1, Action::Move { dx: 1, dy: 0 })],
+            vec![(1, Action::Cast)],
+            vec![(1, Action::Attack { target: 2 })],
+            vec![(1, Action::Wait), (2, Action::Wait)],
+            vec![(1, Action::Cast)],
+        ]
+    }
+
+    #[test]
+    fn match_resolves_through_independent_interpreter() {
+        let mut m = Match::new(MatchId(1), 7, solo_roster("dev"));
+        m.set_program(1, vec![rune(0, 1, 2, 3, 4), rune(5, 2, 1, 0, -3)]);
+        let before = m.world.content_hash();
+        m.resolve_turn(&[(1, Action::Cast)]);
+        // A cast changed the shared world via the independent interpreter.
+        assert_ne!(before, m.world.content_hash());
+        assert_eq!(m.turn, 1);
+    }
+
+    #[test]
+    fn replay_reproduces_final_hash() {
+        let seed = 12345;
+        let roster = solo_roster("dev");
+        let mut m = Match::new(MatchId(seed), seed, roster.clone());
+        m.set_program(1, vec![rune(2, 3, 4, 5, 6), rune(8, 1, 1, 1, 1), rune(0, 7, 7, 7, 7)]);
+        for subs in scripted_match() {
+            m.resolve_turn(&subs);
+        }
+        let original = m.replay.final_hash;
+        // Replay from the recorded inputs alone.
+        let reconstructed = replay(seed, &roster, &m.replay.turns);
+        assert_eq!(original, reconstructed.final_hash, "replay drifted");
+    }
+
+    #[test]
+    fn many_matches_are_deterministic() {
+        for seed in 0..200u64 {
+            let roster = solo_roster("p");
+            let (m, log) = run_scripted(seed, &roster, &scripted_match());
+            let again = replay(seed, &roster, &log.turns);
+            assert_eq!(m.replay.final_hash, again.final_hash, "seed {seed} not deterministic");
+        }
+    }
+
+    #[test]
+    fn hidden_state_never_appears_in_visible_snapshot() {
+        let mut m = Match::new(MatchId(1), 999, solo_roster("dev"));
+        // Mask some observed lanes so redaction is non-trivial.
+        for d in 0..NUM_DOMAINS {
+            m.world.observation_state.visible[d][1] = false;
+        }
+        m.set_program(1, vec![rune(10, 1, 2, 3, 4)]);
+        m.resolve_turn(&[(1, Action::Cast)]);
+        let snap = m.visible_for(1);
+        for vd in &snap.observed_domains {
+            for (l, o) in vd.observed.iter().enumerate() {
+                if !m.world.observation_state.visible[vd.index as usize][l] {
+                    assert!(o.is_none(), "masked lane leaked a value");
+                }
+            }
+        }
+        assert!(snap.hidden_state_redactions >= (NUM_DOMAINS * HIDDEN_LANES) as u32);
+    }
+
+    #[test]
+    fn diagnostics_are_names_and_counts_only() {
+        let m = Match::new(MatchId(1), 5, solo_roster("dev"));
+        let diag = m.diagnostics_for_player(1);
+        for s in diag.known_reads.iter().chain(diag.known_writes.iter()) {
+            assert!(s.chars().any(|c| c.is_alphabetic()), "diagnostic should be a domain name");
+        }
+    }
+}

crates/generators/src/lib.rs

@@ -42,6 +42,10 @@ pub struct PerturbedCase {
     pub axis_name: String,
     pub world: WorldSnapshot,
     pub expectation: TraceDifferenceExpectation,
+    /// The domain this axis perturbs. The metamorphic gate uses this to enforce
+    /// the sound relation "a perturbation the program *consumes* must alter the
+    /// trace" rather than merely "some hash changed".
+    pub target_domain: usize,
 }
 
 /// A complete generated case (per spec).
@@ -156,6 +160,7 @@ pub fn generate_perturbations(world: &WorldSnapshot, seed: u64, count: usize) ->
             axis_name: axis.name(),
             world: axis.apply(world),
             expectation: axis.expected_trace_difference(),
+            target_domain: axis.target().0 as usize,
         });
     }
     out

crates/protocol/Cargo.toml

@@ -0,0 +1,11 @@
+[package]
+name = "protocol"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+world_model = { path = "../world_model" }
+rune_ir = { path = "../rune_ir" }
+trace_model = { path = "../trace_model" }
+reference_runtime = { path = "../reference_runtime" }

crates/protocol/src/json.rs

@@ -0,0 +1,531 @@
+//! A complete hand-rolled JSON value, serializer, and parser (no external
+//! crates). The orchestrator's `ci_reports::json` is write-only; the protocol
+//! needs to *parse* untrusted client packets too, and parsing must be **total**
+//! — any byte sequence yields `Ok` or `Err`, never a panic. That totality is
+//! what lets the server treat a malformed packet as a deterministic
+//! `ValidationReport` rather than a crash.
+
+use std::collections::BTreeMap;
+use std::fmt;
+
+/// A parsed JSON value. Objects use a `BTreeMap` so key order is canonical,
+/// which keeps re-serialization stable and hashable.
+#[derive(Clone, PartialEq, Debug)]
+pub enum Json {
+    Null,
+    Bool(bool),
+    /// All numbers are carried as `f64`; integer accessors round-trip exact
+    /// values within the safe integer range, which is all the protocol uses.
+    Num(f64),
+    Str(String),
+    Arr(Vec<Json>),
+    Obj(BTreeMap<String, Json>),
+}
+
+impl Json {
+    pub fn s(v: impl Into<String>) -> Json {
+        Json::Str(v.into())
+    }
+    pub fn i(v: i64) -> Json {
+        Json::Num(v as f64)
+    }
+    pub fn u(v: u64) -> Json {
+        Json::Num(v as f64)
+    }
+    pub fn obj(fields: Vec<(&str, Json)>) -> Json {
+        let mut m = BTreeMap::new();
+        for (k, v) in fields {
+            m.insert(k.to_string(), v);
+        }
+        Json::Obj(m)
+    }
+
+    // ---- typed accessors (all fallible, none panic) ----
+
+    pub fn get(&self, key: &str) -> Option<&Json> {
+        match self {
+            Json::Obj(m) => m.get(key),
+            _ => None,
+        }
+    }
+    pub fn as_str(&self) -> Option<&str> {
+        match self {
+            Json::Str(s) => Some(s),
+            _ => None,
+        }
+    }
+    pub fn as_f64(&self) -> Option<f64> {
+        match self {
+            Json::Num(n) => Some(*n),
+            _ => None,
+        }
+    }
+    pub fn as_i64(&self) -> Option<i64> {
+        match self {
+            Json::Num(n) if n.is_finite() => Some(*n as i64),
+            _ => None,
+        }
+    }
+    pub fn as_u64(&self) -> Option<u64> {
+        match self {
+            Json::Num(n) if n.is_finite() && *n >= 0.0 => Some(*n as u64),
+            _ => None,
+        }
+    }
+    pub fn as_u8(&self) -> Option<u8> {
+        self.as_u64().and_then(|v| u8::try_from(v).ok())
+    }
+    pub fn as_bool(&self) -> Option<bool> {
+        match self {
+            Json::Bool(b) => Some(*b),
+            _ => None,
+        }
+    }
+    pub fn as_arr(&self) -> Option<&[Json]> {
+        match self {
+            Json::Arr(a) => Some(a),
+            _ => None,
+        }
+    }
+
+    /// Convenience: required field accessors that produce a descriptive error.
+    pub fn field<'a>(&'a self, key: &str) -> Result<&'a Json, JsonError> {
+        self.get(key).ok_or_else(|| JsonError::Field(key.to_string()))
+    }
+    pub fn str_field(&self, key: &str) -> Result<String, JsonError> {
+        self.field(key)?
+            .as_str()
+            .map(|s| s.to_string())
+            .ok_or_else(|| JsonError::Type(key.to_string(), "string"))
+    }
+    pub fn u64_field(&self, key: &str) -> Result<u64, JsonError> {
+        self.field(key)?
+            .as_u64()
+            .ok_or_else(|| JsonError::Type(key.to_string(), "u64"))
+    }
+    pub fn i64_field(&self, key: &str) -> Result<i64, JsonError> {
+        self.field(key)?
+            .as_i64()
+            .ok_or_else(|| JsonError::Type(key.to_string(), "i64"))
+    }
+    pub fn arr_field<'a>(&'a self, key: &str) -> Result<&'a [Json], JsonError> {
+        self.field(key)?
+            .as_arr()
+            .ok_or_else(|| JsonError::Type(key.to_string(), "array"))
+    }
+
+    // ---- serialization ----
+
+    /// Compact canonical serialization (no whitespace). Deterministic because
+    /// object keys are stored sorted.
+    pub fn to_compact(&self) -> String {
+        let mut out = String::new();
+        self.write(&mut out);
+        out
+    }
+
+    fn write(&self, out: &mut String) {
+        match self {
+            Json::Null => out.push_str("null"),
+            Json::Bool(b) => out.push_str(if *b { "true" } else { "false" }),
+            Json::Num(n) => {
+                if !n.is_finite() {
+                    out.push_str("null");
+                } else if *n == n.trunc() && n.abs() < 9_007_199_254_740_992.0 {
+                    // Exact integer: print without a decimal point.
+                    out.push_str(&(*n as i64).to_string());
+                } else {
+                    out.push_str(&format!("{}", n));
+                }
+            }
+            Json::Str(s) => write_str(out, s),
+            Json::Arr(items) => {
+                out.push('[');
+                for (i, it) in items.iter().enumerate() {
+                    if i > 0 {
+                        out.push(',');
+                    }
+                    it.write(out);
+                }
+                out.push(']');
+            }
+            Json::Obj(m) => {
+                out.push('{');
+                for (i, (k, v)) in m.iter().enumerate() {
+                    if i > 0 {
+                        out.push(',');
+                    }
+                    write_str(out, k);
+                    out.push(':');
+                    v.write(out);
+                }
+                out.push('}');
+            }
+        }
+    }
+}
+
+fn write_str(out: &mut String, s: &str) {
+    out.push('"');
+    for c in s.chars() {
+        match c {
+            '"' => out.push_str("\\\""),
+            '\\' => out.push_str("\\\\"),
+            '\n' => out.push_str("\\n"),
+            '\r' => out.push_str("\\r"),
+            '\t' => out.push_str("\\t"),
+            c if (c as u32) < 0x20 => out.push_str(&format!("\\u{:04x}", c as u32)),
+            c => out.push(c),
+        }
+    }
+    out.push('"');
+}
+
+/// A JSON parse / shape error. Carrying a message keeps decode failures
+/// diagnosable without ever unwinding.
+#[derive(Clone, PartialEq, Eq, Debug)]
+pub enum JsonError {
+    Parse(String),
+    Field(String),
+    Type(String, &'static str),
+}
+
+impl fmt::Display for JsonError {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            JsonError::Parse(m) => write!(f, "json parse error: {m}"),
+            JsonError::Field(k) => write!(f, "missing field: {k}"),
+            JsonError::Type(k, t) => write!(f, "field {k} is not a {t}"),
+        }
+    }
+}
+
+impl std::error::Error for JsonError {}
+
+/// Parse a JSON document. Total: never panics on any input.
+pub fn parse(input: &str) -> Result<Json, JsonError> {
+    let bytes = input.as_bytes();
+    let mut p = Parser { bytes, pos: 0, depth: 0 };
+    p.skip_ws();
+    let v = p.value()?;
+    p.skip_ws();
+    if p.pos != bytes.len() {
+        return Err(JsonError::Parse("trailing characters".into()));
+    }
+    Ok(v)
+}
+
+/// Maximum nesting depth. Bounds recursion so a deeply-nested adversarial
+/// packet returns `Err` instead of overflowing the stack.
+const MAX_DEPTH: usize = 64;
+
+struct Parser<'a> {
+    bytes: &'a [u8],
+    pos: usize,
+    depth: usize,
+}
+
+impl<'a> Parser<'a> {
+    fn peek(&self) -> Option<u8> {
+        self.bytes.get(self.pos).copied()
+    }
+
+    fn skip_ws(&mut self) {
+        while let Some(b) = self.peek() {
+            if b == b' ' || b == b'\t' || b == b'\n' || b == b'\r' {
+                self.pos += 1;
+            } else {
+                break;
+            }
+        }
+    }
+
+    fn value(&mut self) -> Result<Json, JsonError> {
+        self.depth += 1;
+        if self.depth > MAX_DEPTH {
+            return Err(JsonError::Parse("max depth exceeded".into()));
+        }
+        let r = match self.peek() {
+            Some(b'{') => self.object(),
+            Some(b'[') => self.array(),
+            Some(b'"') => Ok(Json::Str(self.string()?)),
+            Some(b't') | Some(b'f') => self.boolean(),
+            Some(b'n') => self.null(),
+            Some(b'-') | Some(b'0'..=b'9') => self.number(),
+            Some(c) => Err(JsonError::Parse(format!("unexpected byte '{}'", c as char))),
+            None => Err(JsonError::Parse("unexpected end".into())),
+        };
+        self.depth -= 1;
+        r
+    }
+
+    fn expect(&mut self, b: u8) -> Result<(), JsonError> {
+        if self.peek() == Some(b) {
+            self.pos += 1;
+            Ok(())
+        } else {
+            Err(JsonError::Parse(format!("expected '{}'", b as char)))
+        }
+    }
+
+    fn object(&mut self) -> Result<Json, JsonError> {
+        self.expect(b'{')?;
+        let mut m = BTreeMap::new();
+        self.skip_ws();
+        if self.peek() == Some(b'}') {
+            self.pos += 1;
+            return Ok(Json::Obj(m));
+        }
+        loop {
+            self.skip_ws();
+            let key = self.string()?;
+            self.skip_ws();
+            self.expect(b':')?;
+            self.skip_ws();
+            let val = self.value()?;
+            m.insert(key, val);
+            self.skip_ws();
+            match self.peek() {
+                Some(b',') => {
+                    self.pos += 1;
+                    continue;
+                }
+                Some(b'}') => {
+                    self.pos += 1;
+                    break;
+                }
+                _ => return Err(JsonError::Parse("expected ',' or '}'".into())),
+            }
+        }
+        Ok(Json::Obj(m))
+    }
+
+    fn array(&mut self) -> Result<Json, JsonError> {
+        self.expect(b'[')?;
+        let mut a = Vec::new();
+        self.skip_ws();
+        if self.peek() == Some(b']') {
+            self.pos += 1;
+            return Ok(Json::Arr(a));
+        }
+        loop {
+            self.skip_ws();
+            a.push(self.value()?);
+            self.skip_ws();
+            match self.peek() {
+                Some(b',') => {
+                    self.pos += 1;
+                    continue;
+                }
+                Some(b']') => {
+                    self.pos += 1;
+                    break;
+                }
+                _ => return Err(JsonError::Parse("expected ',' or ']'".into())),
+            }
+        }
+        Ok(Json::Arr(a))
+    }
+
+    fn string(&mut self) -> Result<String, JsonError> {
+        self.expect(b'"')?;
+        let mut s = String::new();
+        loop {
+            match self.peek() {
+                None => return Err(JsonError::Parse("unterminated string".into())),
+                Some(b'"') => {
+                    self.pos += 1;
+                    break;
+                }
+                Some(b'\\') => {
+                    self.pos += 1;
+                    match self.peek() {
+                        Some(b'"') => s.push('"'),
+                        Some(b'\\') => s.push('\\'),
+                        Some(b'/') => s.push('/'),
+                        Some(b'n') => s.push('\n'),
+                        Some(b'r') => s.push('\r'),
+                        Some(b't') => s.push('\t'),
+                        Some(b'b') => s.push('\u{0008}'),
+                        Some(b'f') => s.push('\u{000c}'),
+                        Some(b'u') => {
+                            let cp = self.hex4()?;
+                            // Handle surrogate pairs.
+                            if (0xD800..=0xDBFF).contains(&cp) {
+                                if self.peek() == Some(b'\\') {
+                                    self.pos += 1;
+                                    if self.peek() == Some(b'u') {
+                                        let lo = self.hex4()?;
+                                        if (0xDC00..=0xDFFF).contains(&lo) {
+                                            let c = 0x10000
+                                                + ((cp - 0xD800) << 10)
+                                                + (lo - 0xDC00);
+                                            if let Some(ch) = char::from_u32(c) {
+                                                s.push(ch);
+                                            } else {
+                                                s.push('\u{FFFD}');
+                                            }
+                                            continue;
+                                        }
+                                    }
+                                }
+                                s.push('\u{FFFD}');
+                            } else if let Some(ch) = char::from_u32(cp) {
+                                s.push(ch);
+                            } else {
+                                s.push('\u{FFFD}');
+                            }
+                            continue;
+                        }
+                        _ => return Err(JsonError::Parse("bad escape".into())),
+                    }
+                    self.pos += 1;
+                }
+                Some(_) => {
+                    // Copy one UTF-8 codepoint from the source.
+                    let start = self.pos;
+                    let len = utf8_len(self.bytes[start]);
+                    if start + len > self.bytes.len() {
+                        return Err(JsonError::Parse("bad utf8".into()));
+                    }
+                    match std::str::from_utf8(&self.bytes[start..start + len]) {
+                        Ok(chunk) => s.push_str(chunk),
+                        Err(_) => return Err(JsonError::Parse("bad utf8".into())),
+                    }
+                    self.pos += len;
+                }
+            }
+        }
+        Ok(s)
+    }
+
+    fn hex4(&mut self) -> Result<u32, JsonError> {
+        // assumes the 'u' has been consumed
+        self.pos += 1;
+        let mut v: u32 = 0;
+        for _ in 0..4 {
+            let d = self
+                .peek()
+                .and_then(|b| (b as char).to_digit(16))
+                .ok_or_else(|| JsonError::Parse("bad \\u".into()))?;
+            v = v * 16 + d;
+            self.pos += 1;
+        }
+        Ok(v)
+    }
+
+    fn boolean(&mut self) -> Result<Json, JsonError> {
+        if self.bytes[self.pos..].starts_with(b"true") {
+            self.pos += 4;
+            Ok(Json::Bool(true))
+        } else if self.bytes[self.pos..].starts_with(b"false") {
+            self.pos += 5;
+            Ok(Json::Bool(false))
+        } else {
+            Err(JsonError::Parse("bad literal".into()))
+        }
+    }
+
+    fn null(&mut self) -> Result<Json, JsonError> {
+        if self.bytes[self.pos..].starts_with(b"null") {
+            self.pos += 4;
+            Ok(Json::Null)
+        } else {
+            Err(JsonError::Parse("bad literal".into()))
+        }
+    }
+
+    fn number(&mut self) -> Result<Json, JsonError> {
+        let start = self.pos;
+        if self.peek() == Some(b'-') {
+            self.pos += 1;
+        }
+        while let Some(b'0'..=b'9') = self.peek() {
+            self.pos += 1;
+        }
+        if self.peek() == Some(b'.') {
+            self.pos += 1;
+            while let Some(b'0'..=b'9') = self.peek() {
+                self.pos += 1;
+            }
+        }
+        if let Some(b'e') | Some(b'E') = self.peek() {
+            self.pos += 1;
+            if let Some(b'+') | Some(b'-') = self.peek() {
+                self.pos += 1;
+            }
+            while let Some(b'0'..=b'9') = self.peek() {
+                self.pos += 1;
+            }
+        }
+        let slice = std::str::from_utf8(&self.bytes[start..self.pos])
+            .map_err(|_| JsonError::Parse("bad number".into()))?;
+        slice
+            .parse::<f64>()
+            .map(Json::Num)
+            .map_err(|_| JsonError::Parse("bad number".into()))
+    }
+}
+
+fn utf8_len(b: u8) -> usize {
+    if b < 0x80 {
+        1
+    } else if b >> 5 == 0b110 {
+        2
+    } else if b >> 4 == 0b1110 {
+        3
+    } else if b >> 3 == 0b11110 {
+        4
+    } else {
+        1
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn roundtrip_basic() {
+        let v = Json::obj(vec![
+            ("a", Json::i(42)),
+            ("b", Json::Arr(vec![Json::Bool(true), Json::Null, Json::s("x")])),
+            ("c", Json::Num(1.5)),
+        ]);
+        let s = v.to_compact();
+        let back = parse(&s).unwrap();
+        assert_eq!(v, back);
+    }
+
+    #[test]
+    fn parse_never_panics_on_garbage() {
+        let deep = "{".repeat(1000);
+        let cases = [
+            "", "{", "[", "\"", "nul", "{\"a\":}", "[1,2,", "tru", "12.3.4",
+            "{\"a\"1}", "\\", "\"\\u00\"", deep.as_str(),
+        ];
+        for c in cases {
+            // Must return without panicking; value is irrelevant.
+            let _ = parse(c);
+        }
+    }
+
+    #[test]
+    fn deep_nesting_is_rejected_not_overflowed() {
+        let deep = "[".repeat(10_000);
+        assert!(parse(&deep).is_err());
+    }
+
+    #[test]
+    fn integers_roundtrip_exact() {
+        let v = Json::i(-1234567890123);
+        assert_eq!(parse(&v.to_compact()).unwrap().as_i64(), Some(-1234567890123));
+    }
+
+    #[test]
+    fn escapes_roundtrip() {
+        let v = Json::s("line\ntab\tquote\"slash\\end");
+        let s = v.to_compact();
+        assert_eq!(parse(&s).unwrap(), v);
+    }
+}

crates/protocol/src/lib.rs

@@ -0,0 +1,858 @@
+//! `protocol` — the versioned, serializable, hashable client/server message
+//! contract (Phase A of `plan2.md`). Defined **before** any UI.
+//!
+//! Invariants enforced here:
+//! * Every message carries a protocol version (`PROTOCOL_VERSION`); a decoder
+//!   rejects mismatched versions deterministically.
+//! * Decoding is **total**: any byte string yields `Ok(msg)` or `Err(..)`,
+//!   never a panic. The server relies on this to turn a malformed client
+//!   packet into a `ValidationReport`/`ErrorEvent` instead of crashing.
+//! * Every server output is **hashable** ([`ServerMessage::content_hash`]) over
+//!   a canonical (sorted-key, whitespace-free) serialization, so replays and
+//!   the browser can verify byte-for-byte agreement with the server.
+//! * No game truth lives client-side: client messages carry only *intent*
+//!   (movement choice, rune program, slot selection, inspection request).
+
+pub mod json;
+
+pub use json::{parse, Json, JsonError};
+use world_model::{Hash, Hasher};
+
+/// Protocol version. Bumped on any wire-incompatible change. Both peers check
+/// it on every message.
+pub const PROTOCOL_VERSION: u32 = 1;
+
+/// Stable identifier for a connected player within a match.
+#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
+pub struct PlayerId(pub u32);
+
+/// Stable identifier for a match.
+#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
+pub struct MatchId(pub u64);
+
+// ---------------------------------------------------------------------------
+// Rune token wire form (mirror of `rune_ir::RuneToken`, kept independent so the
+// wire format does not silently change when the IR changes).
+// ---------------------------------------------------------------------------
+
+/// One rune token as it crosses the wire. `op` is the opcode index
+/// (`rune_ir::Op::to_u8`); every field is interpreted modulo its range by the
+/// runtime, so no token value is ever rejected.
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+pub struct RuneTokenWire {
+    pub op: u8,
+    pub a: u8,
+    pub b: u8,
+    pub c: u8,
+    pub imm: i64,
+}
+
+impl RuneTokenWire {
+    pub fn to_json(self) -> Json {
+        Json::obj(vec![
+            ("op", Json::u(self.op as u64)),
+            ("a", Json::u(self.a as u64)),
+            ("b", Json::u(self.b as u64)),
+            ("c", Json::u(self.c as u64)),
+            ("imm", Json::i(self.imm)),
+        ])
+    }
+    pub fn from_json(j: &Json) -> Result<Self, JsonError> {
+        Ok(RuneTokenWire {
+            op: j.field("op")?.as_u8().ok_or(JsonError::Type("op".into(), "u8"))?,
+            a: j.field("a")?.as_u8().ok_or(JsonError::Type("a".into(), "u8"))?,
+            b: j.field("b")?.as_u8().ok_or(JsonError::Type("b".into(), "u8"))?,
+            c: j.field("c")?.as_u8().ok_or(JsonError::Type("c".into(), "u8"))?,
+            imm: j.i64_field("imm")?,
+        })
+    }
+    pub fn into_token(self) -> rune_ir::RuneToken {
+        rune_ir::RuneToken {
+            op: rune_ir::Op::from_u8(self.op),
+            a: self.a,
+            b: self.b,
+            c: self.c,
+            imm: self.imm,
+        }
+    }
+    pub fn from_token(t: &rune_ir::RuneToken) -> Self {
+        RuneTokenWire { op: t.op.to_u8(), a: t.a, b: t.b, c: t.c, imm: t.imm }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Player intent / actions (client -> server only).
+// ---------------------------------------------------------------------------
+
+/// A single per-turn action chosen by a player. The browser only ever sends
+/// *intent*; the server is the sole authority on the outcome.
+#[derive(Clone, PartialEq, Debug)]
+pub enum Action {
+    /// Step one cell on the arena grid (`dx`,`dy` in {-1,0,1}).
+    Move { dx: i32, dy: i32 },
+    /// Inspect a target entity (request diagnostics about it).
+    Inspect { target: u32 },
+    /// Cast the player's currently-edited rune program.
+    Cast,
+    /// Basic stick/melee attack against a target entity.
+    Attack { target: u32 },
+    /// Pass the turn.
+    Wait,
+}
+
+impl Action {
+    pub fn to_json(&self) -> Json {
+        match self {
+            Action::Move { dx, dy } => Json::obj(vec![
+                ("kind", Json::s("move")),
+                ("dx", Json::i(*dx as i64)),
+                ("dy", Json::i(*dy as i64)),
+            ]),
+            Action::Inspect { target } => Json::obj(vec![
+                ("kind", Json::s("inspect")),
+                ("target", Json::u(*target as u64)),
+            ]),
+            Action::Cast => Json::obj(vec![("kind", Json::s("cast"))]),
+            Action::Attack { target } => Json::obj(vec![
+                ("kind", Json::s("attack")),
+                ("target", Json::u(*target as u64)),
+            ]),
+            Action::Wait => Json::obj(vec![("kind", Json::s("wait"))]),
+        }
+    }
+    pub fn from_json(j: &Json) -> Result<Self, JsonError> {
+        match j.str_field("kind")?.as_str() {
+            "move" => {
+                let dx = j.i64_field("dx")? as i32;
+                let dy = j.i64_field("dy")? as i32;
+                // Clamp to legal step range so a hostile client cannot teleport.
+                Ok(Action::Move { dx: dx.clamp(-1, 1), dy: dy.clamp(-1, 1) })
+            }
+            "inspect" => Ok(Action::Inspect { target: j.u64_field("target")? as u32 }),
+            "cast" => Ok(Action::Cast),
+            "attack" => Ok(Action::Attack { target: j.u64_field("target")? as u32 }),
+            "wait" => Ok(Action::Wait),
+            other => Err(JsonError::Parse(format!("unknown action kind '{other}'"))),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// ClientMessage.
+// ---------------------------------------------------------------------------
+
+/// Everything a browser may send. Intent only — never game truth.
+#[derive(Clone, PartialEq, Debug)]
+pub enum ClientMessage {
+    /// Request to join (or create) a match. `name` is a dev/anonymous label.
+    JoinMatch { name: String, match_id: Option<MatchId> },
+    /// Submit this turn's action for the current turn number.
+    SubmitTurn { turn: u64, action: Action },
+    /// Replace the player's editable rune program (library/editor state).
+    EditRuneProgram { tokens: Vec<RuneTokenWire> },
+    /// Ask for diagnostics about a target entity.
+    InspectTarget { target: u32 },
+    /// Ask the server to stream the recorded replay for a match.
+    RequestReplay { match_id: MatchId },
+    /// Liveness ping.
+    Ping { nonce: u64 },
+}
+
+impl ClientMessage {
+    fn type_tag(&self) -> &'static str {
+        match self {
+            ClientMessage::JoinMatch { .. } => "JoinMatch",
+            ClientMessage::SubmitTurn { .. } => "SubmitTurn",
+            ClientMessage::EditRuneProgram { .. } => "EditRuneProgram",
+            ClientMessage::InspectTarget { .. } => "InspectTarget",
+            ClientMessage::RequestReplay { .. } => "RequestReplay",
+            ClientMessage::Ping { .. } => "Ping",
+        }
+    }
+
+    fn body(&self) -> Json {
+        match self {
+            ClientMessage::JoinMatch { name, match_id } => Json::obj(vec![
+                ("name", Json::s(name.clone())),
+                (
+                    "match_id",
+                    match match_id {
+                        Some(m) => Json::u(m.0),
+                        None => Json::Null,
+                    },
+                ),
+            ]),
+            ClientMessage::SubmitTurn { turn, action } => Json::obj(vec![
+                ("turn", Json::u(*turn)),
+                ("action", action.to_json()),
+            ]),
+            ClientMessage::EditRuneProgram { tokens } => Json::obj(vec![(
+                "tokens",
+                Json::Arr(tokens.iter().map(|t| t.to_json()).collect()),
+            )]),
+            ClientMessage::InspectTarget { target } => {
+                Json::obj(vec![("target", Json::u(*target as u64))])
+            }
+            ClientMessage::RequestReplay { match_id } => {
+                Json::obj(vec![("match_id", Json::u(match_id.0))])
+            }
+            ClientMessage::Ping { nonce } => Json::obj(vec![("nonce", Json::u(*nonce))]),
+        }
+    }
+
+    /// Canonical envelope: `{v, type, body}`.
+    pub fn to_json(&self) -> Json {
+        Json::obj(vec![
+            ("v", Json::u(PROTOCOL_VERSION as u64)),
+            ("type", Json::s(self.type_tag())),
+            ("body", self.body()),
+        ])
+    }
+
+    pub fn encode(&self) -> String {
+        self.to_json().to_compact()
+    }
+
+    /// Decode a wire string. Total: never panics. Rejects version mismatch.
+    pub fn decode(raw: &str) -> Result<ClientMessage, JsonError> {
+        let j = parse(raw)?;
+        Self::from_json(&j)
+    }
+
+    pub fn from_json(j: &Json) -> Result<ClientMessage, JsonError> {
+        let v = j.u64_field("v")?;
+        if v != PROTOCOL_VERSION as u64 {
+            return Err(JsonError::Parse(format!(
+                "protocol version mismatch: got {v}, expected {PROTOCOL_VERSION}"
+            )));
+        }
+        let ty = j.str_field("type")?;
+        let body = j.field("body")?;
+        match ty.as_str() {
+            "JoinMatch" => {
+                let name = body.str_field("name")?;
+                let match_id = match body.get("match_id") {
+                    Some(Json::Null) | None => None,
+                    Some(other) => other.as_u64().map(MatchId),
+                };
+                Ok(ClientMessage::JoinMatch { name, match_id })
+            }
+            "SubmitTurn" => {
+                let turn = body.u64_field("turn")?;
+                let action = Action::from_json(body.field("action")?)?;
+                Ok(ClientMessage::SubmitTurn { turn, action })
+            }
+            "EditRuneProgram" => {
+                let arr = body.arr_field("tokens")?;
+                // Bound the program length defensively.
+                if arr.len() > MAX_PROGRAM_TOKENS {
+                    return Err(JsonError::Parse("program too long".into()));
+                }
+                let mut tokens = Vec::with_capacity(arr.len());
+                for t in arr {
+                    tokens.push(RuneTokenWire::from_json(t)?);
+                }
+                Ok(ClientMessage::EditRuneProgram { tokens })
+            }
+            "InspectTarget" => Ok(ClientMessage::InspectTarget {
+                target: body.u64_field("target")? as u32,
+            }),
+            "RequestReplay" => Ok(ClientMessage::RequestReplay {
+                match_id: MatchId(body.u64_field("match_id")?),
+            }),
+            "Ping" => Ok(ClientMessage::Ping { nonce: body.u64_field("nonce")? }),
+            other => Err(JsonError::Parse(format!("unknown client message '{other}'"))),
+        }
+    }
+}
+
+/// Hard upper bound on a submitted rune program, enforced at decode.
+pub const MAX_PROGRAM_TOKENS: usize = 256;
+
+// ---------------------------------------------------------------------------
+// Visibility / knowledge layer (Phase F). This is filtered *game state*, not UI
+// notes: the client renders exactly what the server says is observable, and the
+// hidden ground truth never crosses the wire.
+// ---------------------------------------------------------------------------
+
+/// How well a piece of state is known to the observing player.
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+pub enum Knowledge {
+    Known,
+    Unknown,
+    Suspected,
+    Contradicted,
+    NewlyObserved,
+}
+
+impl Knowledge {
+    pub fn name(self) -> &'static str {
+        match self {
+            Knowledge::Known => "known",
+            Knowledge::Unknown => "unknown",
+            Knowledge::Suspected => "suspected",
+            Knowledge::Contradicted => "contradicted",
+            Knowledge::NewlyObserved => "newly_observed",
+        }
+    }
+    pub fn from_str(s: &str) -> Option<Knowledge> {
+        Some(match s {
+            "known" => Knowledge::Known,
+            "unknown" => Knowledge::Unknown,
+            "suspected" => Knowledge::Suspected,
+            "contradicted" => Knowledge::Contradicted,
+            "newly_observed" => Knowledge::NewlyObserved,
+            _ => return None,
+        })
+    }
+}
+
+/// One domain as the player observes it. Only *visible observed* lanes carry a
+/// value; non-visible observed lanes and **all hidden lanes** are redacted.
+#[derive(Clone, PartialEq, Debug)]
+pub struct VisibleDomain {
+    pub index: u8,
+    pub name: String,
+    /// `Some(v)` for a visible observed lane, `None` for a redacted lane.
+    pub observed: Vec<Option<i64>>,
+    /// Per-lane knowledge tag.
+    pub knowledge: Vec<Knowledge>,
+}
+
+impl VisibleDomain {
+    fn to_json(&self) -> Json {
+        Json::obj(vec![
+            ("index", Json::u(self.index as u64)),
+            ("name", Json::s(self.name.clone())),
+            (
+                "observed",
+                Json::Arr(
+                    self.observed
+                        .iter()
+                        .map(|o| match o {
+                            Some(v) => Json::i(*v),
+                            None => Json::Null,
+                        })
+                        .collect(),
+                ),
+            ),
+            (
+                "knowledge",
+                Json::Arr(self.knowledge.iter().map(|k| Json::s(k.name())).collect()),
+            ),
+        ])
+    }
+    fn from_json(j: &Json) -> Result<Self, JsonError> {
+        let observed = j
+            .arr_field("observed")?
+            .iter()
+            .map(|v| match v {
+                Json::Null => None,
+                other => other.as_i64(),
+            })
+            .collect();
+        let knowledge = j
+            .arr_field("knowledge")?
+            .iter()
+            .filter_map(|v| v.as_str().and_then(Knowledge::from_str))
+            .collect();
+        Ok(VisibleDomain {
+            index: j.field("index")?.as_u8().ok_or(JsonError::Type("index".into(), "u8"))?,
+            name: j.str_field("name")?,
+            observed,
+            knowledge,
+        })
+    }
+}
+
+/// An entity (player/dummy) as seen on the arena.
+#[derive(Clone, PartialEq, Debug)]
+pub struct VisibleEntity {
+    pub id: u32,
+    pub name: String,
+    pub x: i32,
+    pub y: i32,
+    pub hp: i32,
+    pub is_self: bool,
+    pub alive: bool,
+}
+
+impl VisibleEntity {
+    fn to_json(&self) -> Json {
+        Json::obj(vec![
+            ("id", Json::u(self.id as u64)),
+            ("name", Json::s(self.name.clone())),
+            ("x", Json::i(self.x as i64)),
+            ("y", Json::i(self.y as i64)),
+            ("hp", Json::i(self.hp as i64)),
+            ("is_self", Json::Bool(self.is_self)),
+            ("alive", Json::Bool(self.alive)),
+        ])
+    }
+    fn from_json(j: &Json) -> Result<Self, JsonError> {
+        Ok(VisibleEntity {
+            id: j.u64_field("id")? as u32,
+            name: j.str_field("name")?,
+            x: j.i64_field("x")? as i32,
+            y: j.i64_field("y")? as i32,
+            hp: j.i64_field("hp")? as i32,
+            is_self: j.field("is_self")?.as_bool().unwrap_or(false),
+            alive: j.field("alive")?.as_bool().unwrap_or(true),
+        })
+    }
+}
+
+/// The server's filtered view of the world for one player (Phase F).
+#[derive(Clone, PartialEq, Debug)]
+pub struct VisibleWorldSnapshot {
+    pub turn: u64,
+    pub arena_w: i32,
+    pub arena_h: i32,
+    pub observed_domains: Vec<VisibleDomain>,
+    pub observed_entities: Vec<VisibleEntity>,
+    /// Short human-readable environment descriptors (arena conditions).
+    pub observed_environment: Vec<String>,
+    /// Prior-turn outcome summaries the player has already witnessed.
+    pub known_history: Vec<String>,
+    /// Inferred (suspected) markers, e.g. "domain 3 likely volatile".
+    pub inferred_markers: Vec<String>,
+    /// Count of state values deliberately withheld (hidden lanes + masked
+    /// observed lanes). Proof that hidden state exists and is *not* sent.
+    pub hidden_state_redactions: u32,
+}
+
+impl VisibleWorldSnapshot {
+    pub fn to_json(&self) -> Json {
+        Json::obj(vec![
+            ("turn", Json::u(self.turn)),
+            ("arena_w", Json::i(self.arena_w as i64)),
+            ("arena_h", Json::i(self.arena_h as i64)),
+            (
+                "observed_domains",
+                Json::Arr(self.observed_domains.iter().map(|d| d.to_json()).collect()),
+            ),
+            (
+                "observed_entities",
+                Json::Arr(self.observed_entities.iter().map(|e| e.to_json()).collect()),
+            ),
+            (
+                "observed_environment",
+                Json::Arr(self.observed_environment.iter().map(|s| Json::s(s.clone())).collect()),
+            ),
+            (
+                "known_history",
+                Json::Arr(self.known_history.iter().map(|s| Json::s(s.clone())).collect()),
+            ),
+            (
+                "inferred_markers",
+                Json::Arr(self.inferred_markers.iter().map(|s| Json::s(s.clone())).collect()),
+            ),
+            ("hidden_state_redactions", Json::u(self.hidden_state_redactions as u64)),
+        ])
+    }
+    pub fn from_json(j: &Json) -> Result<Self, JsonError> {
+        let observed_domains = j
+            .arr_field("observed_domains")?
+            .iter()
+            .map(VisibleDomain::from_json)
+            .collect::<Result<_, _>>()?;
+        let observed_entities = j
+            .arr_field("observed_entities")?
+            .iter()
+            .map(VisibleEntity::from_json)
+            .collect::<Result<_, _>>()?;
+        let strs = |key| -> Result<Vec<String>, JsonError> {
+            Ok(j.arr_field(key)?
+                .iter()
+                .filter_map(|v| v.as_str().map(|s| s.to_string()))
+                .collect())
+        };
+        Ok(VisibleWorldSnapshot {
+            turn: j.u64_field("turn")?,
+            arena_w: j.i64_field("arena_w")? as i32,
+            arena_h: j.i64_field("arena_h")? as i32,
+            observed_domains,
+            observed_entities,
+            observed_environment: strs("observed_environment")?,
+            known_history: strs("known_history")?,
+            inferred_markers: strs("inferred_markers")?,
+            hidden_state_redactions: j.u64_field("hidden_state_redactions")? as u32,
+        })
+    }
+}
+
+/// Player-facing diagnostics for a rune program (Phase D). Strictly *observed*
+/// claims — never "guaranteed damage" or full hidden state.
+#[derive(Clone, PartialEq, Debug, Default)]
+pub struct RuneDiagnostics {
+    pub known_reads: Vec<String>,
+    pub known_writes: Vec<String>,
+    pub observed_risks: Vec<String>,
+    pub unknown_listeners: u32,
+    pub previous_outcomes: Vec<String>,
+}
+
+impl RuneDiagnostics {
+    fn to_json(&self) -> Json {
+        Json::obj(vec![
+            ("known_reads", Json::Arr(self.known_reads.iter().map(|s| Json::s(s.clone())).collect())),
+            ("known_writes", Json::Arr(self.known_writes.iter().map(|s| Json::s(s.clone())).collect())),
+            ("observed_risks", Json::Arr(self.observed_risks.iter().map(|s| Json::s(s.clone())).collect())),
+            ("unknown_listeners", Json::u(self.unknown_listeners as u64)),
+            ("previous_outcomes", Json::Arr(self.previous_outcomes.iter().map(|s| Json::s(s.clone())).collect())),
+        ])
+    }
+    fn from_json(j: &Json) -> Result<Self, JsonError> {
+        let strs = |key| -> Vec<String> {
+            j.get(key)
+                .and_then(|v| v.as_arr())
+                .map(|a| a.iter().filter_map(|v| v.as_str().map(|s| s.to_string())).collect())
+                .unwrap_or_default()
+        };
+        Ok(RuneDiagnostics {
+            known_reads: strs("known_reads"),
+            known_writes: strs("known_writes"),
+            observed_risks: strs("observed_risks"),
+            unknown_listeners: j.get("unknown_listeners").and_then(|v| v.as_u64()).unwrap_or(0) as u32,
+            previous_outcomes: strs("previous_outcomes"),
+        })
+    }
+}
+
+/// One recorded turn in a replay stream (Phase G).
+#[derive(Clone, PartialEq, Debug)]
+pub struct ReplayTurn {
+    pub turn: u64,
+    /// `(player_id, action)` pairs applied this turn, in canonical order.
+    pub inputs: Vec<(u32, Action)>,
+    /// The runtime canonical replay hash produced this turn.
+    pub runtime_hash: String,
+}
+
+impl ReplayTurn {
+    fn to_json(&self) -> Json {
+        Json::obj(vec![
+            ("turn", Json::u(self.turn)),
+            (
+                "inputs",
+                Json::Arr(
+                    self.inputs
+                        .iter()
+                        .map(|(pid, a)| {
+                            Json::obj(vec![("player", Json::u(*pid as u64)), ("action", a.to_json())])
+                        })
+                        .collect(),
+                ),
+            ),
+            ("runtime_hash", Json::s(self.runtime_hash.clone())),
+        ])
+    }
+    fn from_json(j: &Json) -> Result<Self, JsonError> {
+        let inputs = j
+            .arr_field("inputs")?
+            .iter()
+            .map(|e| {
+                let pid = e.u64_field("player")? as u32;
+                let a = Action::from_json(e.field("action")?)?;
+                Ok((pid, a))
+            })
+            .collect::<Result<_, JsonError>>()?;
+        Ok(ReplayTurn {
+            turn: j.u64_field("turn")?,
+            inputs,
+            runtime_hash: j.str_field("runtime_hash")?,
+        })
+    }
+}
+
+// ---------------------------------------------------------------------------
+// ServerMessage.
+// ---------------------------------------------------------------------------
+
+/// Everything the server may send. Every variant is hashable; the browser can
+/// verify it matches a recorded replay.
+#[derive(Clone, PartialEq, Debug)]
+pub enum ServerMessage {
+    /// Assigned identity + match parameters on join.
+    MatchState {
+        match_id: MatchId,
+        player_id: PlayerId,
+        turn: u64,
+        snapshot: VisibleWorldSnapshot,
+    },
+    /// A new turn has begun; `deadline_ms` is the wall-clock budget.
+    TurnStarted { turn: u64, deadline_ms: u64 },
+    /// A turn resolved authoritatively. Carries the runtime replay hash so the
+    /// client can verify determinism.
+    TurnResolved {
+        turn: u64,
+        snapshot: VisibleWorldSnapshot,
+        runtime_hash: String,
+        events: Vec<String>,
+    },
+    /// Result of an inspection request (filtered observations of a target).
+    ObservationResult { target: u32, diagnostics: RuneDiagnostics },
+    /// Validation feedback for a client packet (accepted/rejected + why).
+    ValidationReport { accepted: bool, detail: String, diagnostics: RuneDiagnostics },
+    /// One chunk of a replay stream.
+    ReplayChunk {
+        match_id: MatchId,
+        seed: u64,
+        index: u32,
+        total: u32,
+        turns: Vec<ReplayTurn>,
+        final_hash: String,
+    },
+    /// A protocol/transport error that is not tied to a specific submission.
+    ErrorEvent { code: String, detail: String },
+}
+
+impl ServerMessage {
+    fn type_tag(&self) -> &'static str {
+        match self {
+            ServerMessage::MatchState { .. } => "MatchState",
+            ServerMessage::TurnStarted { .. } => "TurnStarted",
+            ServerMessage::TurnResolved { .. } => "TurnResolved",
+            ServerMessage::ObservationResult { .. } => "ObservationResult",
+            ServerMessage::ValidationReport { .. } => "ValidationReport",
+            ServerMessage::ReplayChunk { .. } => "ReplayChunk",
+            ServerMessage::ErrorEvent { .. } => "ErrorEvent",
+        }
+    }
+
+    fn body(&self) -> Json {
+        match self {
+            ServerMessage::MatchState { match_id, player_id, turn, snapshot } => Json::obj(vec![
+                ("match_id", Json::u(match_id.0)),
+                ("player_id", Json::u(player_id.0 as u64)),
+                ("turn", Json::u(*turn)),
+                ("snapshot", snapshot.to_json()),
+            ]),
+            ServerMessage::TurnStarted { turn, deadline_ms } => Json::obj(vec![
+                ("turn", Json::u(*turn)),
+                ("deadline_ms", Json::u(*deadline_ms)),
+            ]),
+            ServerMessage::TurnResolved { turn, snapshot, runtime_hash, events } => Json::obj(vec![
+                ("turn", Json::u(*turn)),
+                ("snapshot", snapshot.to_json()),
+                ("runtime_hash", Json::s(runtime_hash.clone())),
+                ("events", Json::Arr(events.iter().map(|s| Json::s(s.clone())).collect())),
+            ]),
+            ServerMessage::ObservationResult { target, diagnostics } => Json::obj(vec![
+                ("target", Json::u(*target as u64)),
+                ("diagnostics", diagnostics.to_json()),
+            ]),
+            ServerMessage::ValidationReport { accepted, detail, diagnostics } => Json::obj(vec![
+                ("accepted", Json::Bool(*accepted)),
+                ("detail", Json::s(detail.clone())),
+                ("diagnostics", diagnostics.to_json()),
+            ]),
+            ServerMessage::ReplayChunk { match_id, seed, index, total, turns, final_hash } => {
+                Json::obj(vec![
+                    ("match_id", Json::u(match_id.0)),
+                    ("seed", Json::u(*seed)),
+                    ("index", Json::u(*index as u64)),
+                    ("total", Json::u(*total as u64)),
+                    ("turns", Json::Arr(turns.iter().map(|t| t.to_json()).collect())),
+                    ("final_hash", Json::s(final_hash.clone())),
+                ])
+            }
+            ServerMessage::ErrorEvent { code, detail } => Json::obj(vec![
+                ("code", Json::s(code.clone())),
+                ("detail", Json::s(detail.clone())),
+            ]),
+        }
+    }
+
+    pub fn to_json(&self) -> Json {
+        Json::obj(vec![
+            ("v", Json::u(PROTOCOL_VERSION as u64)),
+            ("type", Json::s(self.type_tag())),
+            ("body", self.body()),
+        ])
+    }
+
+    pub fn encode(&self) -> String {
+        self.to_json().to_compact()
+    }
+
+    /// Stable content hash over the canonical serialization. Because object
+    /// keys are sorted and there is no whitespace, identical messages hash
+    /// identically across machines — this is how replays are verified.
+    pub fn content_hash(&self) -> Hash {
+        let mut h = Hasher::new();
+        h.write_tag("server-message");
+        h.write_bytes(self.encode().as_bytes());
+        h.finish()
+    }
+
+    pub fn decode(raw: &str) -> Result<ServerMessage, JsonError> {
+        let j = parse(raw)?;
+        Self::from_json(&j)
+    }
+
+    pub fn from_json(j: &Json) -> Result<ServerMessage, JsonError> {
+        let v = j.u64_field("v")?;
+        if v != PROTOCOL_VERSION as u64 {
+            return Err(JsonError::Parse(format!(
+                "protocol version mismatch: got {v}, expected {PROTOCOL_VERSION}"
+            )));
+        }
+        let ty = j.str_field("type")?;
+        let body = j.field("body")?;
+        match ty.as_str() {
+            "MatchState" => Ok(ServerMessage::MatchState {
+                match_id: MatchId(body.u64_field("match_id")?),
+                player_id: PlayerId(body.u64_field("player_id")? as u32),
+                turn: body.u64_field("turn")?,
+                snapshot: VisibleWorldSnapshot::from_json(body.field("snapshot")?)?,
+            }),
+            "TurnStarted" => Ok(ServerMessage::TurnStarted {
+                turn: body.u64_field("turn")?,
+                deadline_ms: body.u64_field("deadline_ms")?,
+            }),
+            "TurnResolved" => Ok(ServerMessage::TurnResolved {
+                turn: body.u64_field("turn")?,
+                snapshot: VisibleWorldSnapshot::from_json(body.field("snapshot")?)?,
+                runtime_hash: body.str_field("runtime_hash")?,
+                events: body
+                    .arr_field("events")?
+                    .iter()
+                    .filter_map(|v| v.as_str().map(|s| s.to_string()))
+                    .collect(),
+            }),
+            "ObservationResult" => Ok(ServerMessage::ObservationResult {
+                target: body.u64_field("target")? as u32,
+                diagnostics: RuneDiagnostics::from_json(body.field("diagnostics")?)?,
+            }),
+            "ValidationReport" => Ok(ServerMessage::ValidationReport {
+                accepted: body.field("accepted")?.as_bool().unwrap_or(false),
+                detail: body.str_field("detail")?,
+                diagnostics: RuneDiagnostics::from_json(body.field("diagnostics")?)?,
+            }),
+            "ReplayChunk" => Ok(ServerMessage::ReplayChunk {
+                match_id: MatchId(body.u64_field("match_id")?),
+                seed: body.u64_field("seed")?,
+                index: body.u64_field("index")? as u32,
+                total: body.u64_field("total")? as u32,
+                turns: body
+                    .arr_field("turns")?
+                    .iter()
+                    .map(ReplayTurn::from_json)
+                    .collect::<Result<_, _>>()?,
+                final_hash: body.str_field("final_hash")?,
+            }),
+            "ErrorEvent" => Ok(ServerMessage::ErrorEvent {
+                code: body.str_field("code")?,
+                detail: body.str_field("detail")?,
+            }),
+            other => Err(JsonError::Parse(format!("unknown server message '{other}'"))),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn sample_snapshot() -> VisibleWorldSnapshot {
+        VisibleWorldSnapshot {
+            turn: 3,
+            arena_w: 8,
+            arena_h: 8,
+            observed_domains: vec![VisibleDomain {
+                index: 0,
+                name: "aether".into(),
+                observed: vec![Some(1), None, Some(-4), None],
+                knowledge: vec![
+                    Knowledge::Known,
+                    Knowledge::Unknown,
+                    Knowledge::NewlyObserved,
+                    Knowledge::Unknown,
+                ],
+            }],
+            observed_entities: vec![VisibleEntity {
+                id: 1,
+                name: "you".into(),
+                x: 2,
+                y: 3,
+                hp: 30,
+                is_self: true,
+                alive: true,
+            }],
+            observed_environment: vec!["calm".into()],
+            known_history: vec!["turn 2: you moved".into()],
+            inferred_markers: vec!["domain 4 likely volatile".into()],
+            hidden_state_redactions: 18,
+        }
+    }
+
+    #[test]
+    fn client_messages_roundtrip() {
+        let msgs = vec![
+            ClientMessage::JoinMatch { name: "dev".into(), match_id: None },
+            ClientMessage::JoinMatch { name: "dev".into(), match_id: Some(MatchId(9)) },
+            ClientMessage::SubmitTurn { turn: 4, action: Action::Move { dx: 1, dy: -1 } },
+            ClientMessage::SubmitTurn { turn: 4, action: Action::Cast },
+            ClientMessage::SubmitTurn { turn: 4, action: Action::Attack { target: 2 } },
+            ClientMessage::EditRuneProgram {
+                tokens: vec![RuneTokenWire { op: 0, a: 1, b: 2, c: 3, imm: -7 }],
+            },
+            ClientMessage::InspectTarget { target: 5 },
+            ClientMessage::RequestReplay { match_id: MatchId(42) },
+            ClientMessage::Ping { nonce: 123 },
+        ];
+        for m in msgs {
+            let s = m.encode();
+            assert_eq!(ClientMessage::decode(&s).unwrap(), m, "roundtrip failed for {m:?}");
+        }
+    }
+
+    #[test]
+    fn server_messages_roundtrip_and_hash_is_stable() {
+        let msgs = vec![
+            ServerMessage::MatchState {
+                match_id: MatchId(1),
+                player_id: PlayerId(1),
+                turn: 0,
+                snapshot: sample_snapshot(),
+            },
+            ServerMessage::TurnStarted { turn: 1, deadline_ms: 5000 },
+            ServerMessage::TurnResolved {
+                turn: 1,
+                snapshot: sample_snapshot(),
+                runtime_hash: "deadbeefcafef00d".into(),
+                events: vec!["you cast".into(), "dummy took 4".into()],
+            },
+            ServerMessage::ValidationReport {
+                accepted: false,
+                detail: "late".into(),
+                diagnostics: RuneDiagnostics::default(),
+            },
+        ];
+        for m in msgs {
+            let s = m.encode();
+            let back = ServerMessage::decode(&s).unwrap();
+            assert_eq!(back, m);
+            // Hash is a pure function of the canonical bytes.
+            assert_eq!(m.content_hash(), back.content_hash());
+        }
+    }
+
+    #[test]
+    fn version_mismatch_is_rejected() {
+        let mut j = ClientMessage::Ping { nonce: 1 }.to_json();
+        if let Json::Obj(ref mut m) = j {
+            m.insert("v".into(), Json::u(999));
+        }
+        assert!(ClientMessage::from_json(&j).is_err());
+    }
+
+    #[test]
+    fn decode_is_total_on_garbage() {
+        for raw in ["", "{}", "null", "{\"v\":1}", "{\"v\":1,\"type\":\"Nope\",\"body\":{}}"] {
+            // Must be Err, never a panic.
+            assert!(ClientMessage::decode(raw).is_err());
+        }
+    }
+}

crates/replay_corpus/Cargo.toml

@@ -10,6 +10,11 @@ rune_ir = { path = "../rune_ir" }
 trace_model = { path = "../trace_model" }
 generators = { path = "../generators" }
 reference_runtime = { path = "../reference_runtime" }
+runtime_under_test = { path = "../runtime_under_test" }
+
+[dev-dependencies]
+# Enables `buggy_resolve` for the retention negative-control test.
+runtime_under_test = { path = "../runtime_under_test", features = ["negative_controls"] }
 
 [lib]
 path = "src/lib.rs"

crates/replay_corpus/corpus/retained_failures.tsv

@@ -0,0 +1,7 @@
+# magicka-retained-failures v1
+seed	gate	note
+0000000000000001	runtime_equivalence	curated regression guard: independent-runtime equivalence on a causal-edge-bearing case (catches the dropped-edge bug class)
+0000000000000007	runtime_equivalence	curated regression guard: fault-heavy case (saturation/empty-accumulator paths)
+000000000000002a	runtime_equivalence	curated regression guard: temporal/future-dependence case
+00000000000000ff	runtime_equivalence	curated regression guard: high-coupling case
+0000000000000539	runtime_equivalence	curated regression guard: branch-divergence case

crates/replay_corpus/src/bin/freeze.rs

@@ -0,0 +1,27 @@
+//! Freeze the committed replay corpus. Run this only as a deliberate, reviewed
+//! migration when the engine semantics legitimately change.
+//!
+//! ```text
+//! cargo run --release -p replay_corpus --bin freeze -- [count]
+//! ```
+
+fn main() {
+    let count: usize = std::env::args()
+        .nth(1)
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(600);
+    let base = replay_corpus::DEFAULT_BASE_SEED;
+    match replay_corpus::freeze_to_disk(count, base) {
+        Ok(n) => {
+            println!(
+                "froze {} replay cases to {}",
+                n,
+                replay_corpus::corpus_path().display()
+            );
+        }
+        Err(e) => {
+            eprintln!("failed to freeze corpus: {e}");
+            std::process::exit(1);
+        }
+    }
+}

crates/replay_corpus/src/lib.rs

@@ -1,12 +1,26 @@
 //! `replay_corpus` — every case is permanent and must replay bit-for-bit.
-//! A replay case stores the seeds plus the three canonical hashes (trace,
+//!
-//! delta, future). Replaying regenerates the case deterministically from its
+//! The corpus is **persisted to a committed file** (`corpus/replay_corpus.tsv`).
-//! master seed, re-executes the reference runtime, and asserts zero hash drift.
+//! Replay loads the expected hashes from that file — produced by an earlier
+//! `freeze` run — regenerates the case deterministically from its master seed,
+//! re-executes the reference runtime, and asserts zero drift against the stored
+//! expectation. Because the expectation is read from disk rather than recomputed
+//! and compared to itself in the same run, drift is genuinely possible: any
+//! change to the engine that alters a hash makes the committed expectation and
+//! the fresh execution disagree, and CI fails. (Proven by the negative-control
+//! test, which corrupts a stored hash and checks the drift is detected.)
 
 use generators::generate_accepted_case;
 use reference_runtime::{execute, EngineConfig, ResolutionInput};
+use std::path::PathBuf;
 use world_model::Hash;
 
+pub mod retention;
+
+/// Format version of the persisted corpus file. Bump only with a deliberate,
+/// reviewed migration of the committed corpus.
+pub const CORPUS_VERSION: u32 = 1;
+
 /// A permanent replay case (per spec) plus the master seed needed to
 /// regenerate the full case deterministically.
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
@@ -21,9 +35,8 @@ pub struct ReplayCase {
     pub expected_future_hash: Hash,
 }
 
-fn input_for(master_seed: u64) -> (ResolutionInput, u64, u64, u64, u64) {
+pub(crate) fn input_for(master_seed: u64) -> (ResolutionInput, u64, u64, u64, u64) {
-    let (case, accepted_seed) = generate_accepted_case(master_seed);
+    let (case, _accepted_seed) = generate_accepted_case(master_seed);
-    let _ = accepted_seed;
     let input = ResolutionInput {
         world: case.world.clone(),
         program: case.program.clone(),
@@ -40,7 +53,13 @@ fn input_for(master_seed: u64) -> (ResolutionInput, u64, u64, u64, u64) {
     )
 }
 
-/// Build a single replay case from a master seed.
+/// Master seed for the `i`-th corpus case (stable, deterministic).
+pub fn master_seed_for(base_seed: u64, i: usize) -> u64 {
+    base_seed ^ (i as u64).wrapping_mul(0x9e3779b97f4a7c15)
+}
+
+/// Build a single replay case from a master seed using the *current* reference.
+/// Used when freezing the corpus to disk.
 pub fn build_case(master_seed: u64) -> ReplayCase {
     let (input, ws, ps, cs, prs) = input_for(master_seed);
     let r = execute(&EngineConfig::reference(), &input);
@@ -56,13 +75,93 @@ pub fn build_case(master_seed: u64) -> ReplayCase {
     }
 }
 
-/// Build a replay corpus of `n` cases.
+/// Build an in-memory corpus of `n` cases (used by `freeze`).
 pub fn build_corpus(n: usize, base_seed: u64) -> Vec<ReplayCase> {
-    (0..n)
+    (0..n).map(|i| build_case(master_seed_for(base_seed, i))).collect()
-        .map(|i| build_case(base_seed ^ (i as u64).wrapping_mul(0x9e3779b97f4a7c15)))
-        .collect()
 }
 
+// --- Persistence. -----------------------------------------------------------
+
+/// Path to the committed corpus file, anchored to this crate.
+pub fn corpus_path() -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("corpus/replay_corpus.tsv")
+}
+
+/// The default base seed the committed corpus is frozen with.
+pub const DEFAULT_BASE_SEED: u64 = 0x5EED;
+
+/// Serialize a corpus to the on-disk TSV format (with a provenance header).
+pub fn serialize_corpus(corpus: &[ReplayCase]) -> String {
+    let mut s = String::new();
+    s.push_str(&format!("# magicka-replay-corpus v{} cases={}\n", CORPUS_VERSION, corpus.len()));
+    s.push_str("master\tworld\tprogram\tcontract\tperturb\ttrace\tdelta\tfuture\n");
+    for c in corpus {
+        s.push_str(&format!(
+            "{:016x}\t{:016x}\t{:016x}\t{:016x}\t{:016x}\t{:016x}\t{:016x}\t{:016x}\n",
+            c.master_seed,
+            c.world_seed,
+            c.program_seed,
+            c.contract_seed,
+            c.perturbation_seed,
+            c.expected_trace_hash.0,
+            c.expected_delta_hash.0,
+            c.expected_future_hash.0,
+        ));
+    }
+    s
+}
+
+fn parse_hex(s: &str) -> Option<u64> {
+    u64::from_str_radix(s.trim(), 16).ok()
+}
+
+/// Parse a corpus from the on-disk TSV format.
+pub fn parse_corpus(text: &str) -> Vec<ReplayCase> {
+    let mut out = Vec::new();
+    for line in text.lines() {
+        if line.starts_with('#') || line.starts_with("master") || line.trim().is_empty() {
+            continue;
+        }
+        let f: Vec<&str> = line.split('\t').collect();
+        if f.len() != 8 {
+            continue;
+        }
+        let vals: Option<Vec<u64>> = f.iter().map(|x| parse_hex(x)).collect();
+        if let Some(v) = vals {
+            out.push(ReplayCase {
+                master_seed: v[0],
+                world_seed: v[1],
+                program_seed: v[2],
+                contract_seed: v[3],
+                perturbation_seed: v[4],
+                expected_trace_hash: Hash(v[5]),
+                expected_delta_hash: Hash(v[6]),
+                expected_future_hash: Hash(v[7]),
+            });
+        }
+    }
+    out
+}
+
+/// Load the committed corpus from disk.
+pub fn load_persisted_corpus() -> std::io::Result<Vec<ReplayCase>> {
+    let text = std::fs::read_to_string(corpus_path())?;
+    Ok(parse_corpus(&text))
+}
+
+/// Freeze a fresh corpus of `n` cases to the committed file.
+pub fn freeze_to_disk(n: usize, base_seed: u64) -> std::io::Result<usize> {
+    let corpus = build_corpus(n, base_seed);
+    let path = corpus_path();
+    if let Some(parent) = path.parent() {
+        std::fs::create_dir_all(parent)?;
+    }
+    std::fs::write(&path, serialize_corpus(&corpus))?;
+    Ok(corpus.len())
+}
+
+// --- Replay verification (against persisted expectations). ------------------
+
 /// A single replay verification outcome.
 #[derive(Clone, Copy, Debug)]
 pub struct ReplayDrift {
@@ -78,15 +177,16 @@ impl ReplayDrift {
     }
 }
 
-/// Replay one case and check for drift.
+/// Replay one *stored* case: regenerate it and compare a fresh reference run to
-pub fn replay(case: &ReplayCase) -> ReplayDrift {
+/// the expectation read from disk.
-    let (input, ..) = input_for(case.master_seed);
+pub fn replay_against_stored(stored: &ReplayCase) -> ReplayDrift {
+    let (input, ..) = input_for(stored.master_seed);
     let r = execute(&EngineConfig::reference(), &input);
     ReplayDrift {
-        master_seed: case.master_seed,
+        master_seed: stored.master_seed,
-        trace_ok: r.trace.canonical_hash() == case.expected_trace_hash,
+        trace_ok: r.trace.canonical_hash() == stored.expected_trace_hash,
-        delta_ok: r.delta.hash() == case.expected_delta_hash,
+        delta_ok: r.delta.hash() == stored.expected_delta_hash,
-        future_ok: r.replay.future_hash == case.expected_future_hash,
+        future_ok: r.replay.future_hash == stored.expected_future_hash,
     }
 }
 
@@ -96,20 +196,21 @@ pub struct ReplayReport {
     pub total: usize,
     pub deterministic: usize,
     pub drift: Vec<u64>,
+    pub loaded_from_disk: bool,
 }
 
 impl ReplayReport {
     pub fn passed(&self, minimum: usize) -> bool {
-        self.drift.is_empty() && self.total >= minimum
+        self.loaded_from_disk && self.drift.is_empty() && self.total >= minimum
     }
 }
 
-/// Verify the whole corpus replays deterministically.
+/// Verify a corpus (already loaded from disk) replays without drift.
 pub fn verify_corpus(corpus: &[ReplayCase]) -> ReplayReport {
     let mut drift = Vec::new();
     let mut deterministic = 0;
     for case in corpus {
-        let d = replay(case);
+        let d = replay_against_stored(case);
         if d.ok() {
             deterministic += 1;
         } else {
@@ -120,6 +221,20 @@ pub fn verify_corpus(corpus: &[ReplayCase]) -> ReplayReport {
         total: corpus.len(),
         deterministic,
         drift,
+        loaded_from_disk: true,
+    }
+}
+
+/// Load the committed corpus and verify it. The single entry point CI uses.
+pub fn verify_persisted_corpus() -> ReplayReport {
+    match load_persisted_corpus() {
+        Ok(corpus) => verify_corpus(&corpus),
+        Err(_) => ReplayReport {
+            total: 0,
+            deterministic: 0,
+            drift: Vec::new(),
+            loaded_from_disk: false,
+        },
     }
 }
 
@@ -128,18 +243,35 @@ mod tests {
     use super::*;
 
     #[test]
-    fn replay_is_deterministic_zero_drift() {
+    fn committed_corpus_loads_and_replays_without_drift() {
-        let corpus = build_corpus(80, 0x5EED);
+        let corpus = load_persisted_corpus().expect("committed corpus must exist; run `freeze`");
+        assert!(!corpus.is_empty(), "committed corpus is empty");
         let report = verify_corpus(&corpus);
-        assert_eq!(report.total, 80);
+        assert!(report.drift.is_empty(), "drift in committed corpus: {:?}", report.drift);
-        assert_eq!(report.deterministic, 80);
+        assert_eq!(report.deterministic, report.total);
-        assert!(report.drift.is_empty());
     }
 
     #[test]
-    fn case_hashes_are_stable() {
+    fn serialize_roundtrips() {
-        let a = build_case(123);
+        let corpus = build_corpus(20, 0x1234);
-        let b = build_case(123);
+        let text = serialize_corpus(&corpus);
-        assert_eq!(a, b);
+        let parsed = parse_corpus(&text);
+        assert_eq!(corpus, parsed);
+    }
+
+    /// Negative control: a corrupted stored expectation is detected as drift.
+    /// Proves the replay gate is not vacuous.
+    #[test]
+    fn corrupted_expectation_is_detected() {
+        let mut corpus = build_corpus(10, 0x9999);
+        // Flip one stored hash — as if the committed corpus disagreed with the
+        // engine. Replay must flag it.
+        corpus[3].expected_trace_hash = Hash(corpus[3].expected_trace_hash.0 ^ 0xdead_beef);
+        let report = verify_corpus(&corpus);
+        assert!(
+            !report.drift.is_empty(),
+            "replay failed to detect a corrupted expectation"
+        );
+        assert!(report.drift.contains(&corpus[3].master_seed));
     }
 }

crates/replay_corpus/src/retention.rs

@@ -0,0 +1,162 @@
+//! Failure retention (finding 6). A static seed corpus proves the engine is
+//! stable on a *fixed* sample; it does not retain *discovered* counterexamples.
+//! This module is the regression memory: every counterexample CI ever finds
+//! (e.g. a reference/runtime divergence) is curated into a committed,
+//! append-only file and **re-verified on every run**, so a fixed bug can never
+//! silently reappear.
+//!
+//! Two halves, both real:
+//! * The committed `corpus/retained_failures.tsv` is loaded and each case is
+//!   re-executed under the reference and the independent runtime-under-test; any
+//!   case where they disagree is a *regression* and fails CI.
+//! * When a run discovers a NEW divergence, it is serialized to the run's output
+//!   so it must be triaged and added to the committed set (the run also fails).
+//!   Discovered failures are therefore never lost.
+
+use crate::input_for;
+use reference_runtime::{canonical, execute, EngineConfig, ResolutionInput, ResolutionResult};
+use runtime_under_test::native_resolve;
+use std::path::PathBuf;
+
+/// A retained counterexample: the seed needed to regenerate it, the gate it
+/// originally tripped, and a human note.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct RetainedFailure {
+    pub master_seed: u64,
+    pub gate: String,
+    pub note: String,
+}
+
+/// Path to the committed retained-failures file.
+pub fn retained_path() -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("corpus/retained_failures.tsv")
+}
+
+/// Serialize a retained set (with a provenance header).
+pub fn serialize(failures: &[RetainedFailure]) -> String {
+    let mut s = String::from("# magicka-retained-failures v1\n");
+    s.push_str("seed\tgate\tnote\n");
+    for f in failures {
+        // Tabs/newlines are stripped from free text to keep the TSV well-formed.
+        let gate = f.gate.replace(['\t', '\n'], " ");
+        let note = f.note.replace(['\t', '\n'], " ");
+        s.push_str(&format!("{:016x}\t{}\t{}\n", f.master_seed, gate, note));
+    }
+    s
+}
+
+/// Parse a retained set.
+pub fn parse(text: &str) -> Vec<RetainedFailure> {
+    let mut out = Vec::new();
+    for line in text.lines() {
+        if line.starts_with('#') || line.starts_with("seed") || line.trim().is_empty() {
+            continue;
+        }
+        let f: Vec<&str> = line.splitn(3, '\t').collect();
+        if f.len() < 1 {
+            continue;
+        }
+        if let Ok(seed) = u64::from_str_radix(f[0].trim(), 16) {
+            out.push(RetainedFailure {
+                master_seed: seed,
+                gate: f.get(1).unwrap_or(&"").to_string(),
+                note: f.get(2).unwrap_or(&"").to_string(),
+            });
+        }
+    }
+    out
+}
+
+/// Load the committed retained set. Missing file => empty set + `present=false`.
+pub fn load() -> (Vec<RetainedFailure>, bool) {
+    match std::fs::read_to_string(retained_path()) {
+        Ok(t) => (parse(&t), true),
+        Err(_) => (Vec::new(), false),
+    }
+}
+
+/// Append a newly-discovered failure to the committed set (append-only).
+pub fn append(failure: &RetainedFailure) -> std::io::Result<()> {
+    let (mut set, _present) = load();
+    if !set.iter().any(|f| f.master_seed == failure.master_seed) {
+        set.push(failure.clone());
+    }
+    let path = retained_path();
+    if let Some(parent) = path.parent() {
+        std::fs::create_dir_all(parent)?;
+    }
+    std::fs::write(&path, serialize(&set))
+}
+
+/// Does a retained case still agree between the reference and `resolve`? A
+/// retained counterexample is "fixed" iff the two implementations now produce
+/// identical canonical views on it.
+pub fn agrees(
+    failure: &RetainedFailure,
+    resolve: impl Fn(&ResolutionInput) -> ResolutionResult,
+) -> bool {
+    let (input, ..) = input_for(failure.master_seed);
+    canonical(&execute(&EngineConfig::reference(), &input)) == canonical(&resolve(&input))
+}
+
+/// Result of verifying the retained set.
+#[derive(Clone, Debug)]
+pub struct RetentionReport {
+    pub present: bool,
+    pub total: usize,
+    /// Seeds that regressed (reference and runtime-under-test disagree again).
+    pub regressions: Vec<u64>,
+}
+
+impl RetentionReport {
+    pub fn ok(&self) -> bool {
+        self.present && self.regressions.is_empty()
+    }
+}
+
+/// Verify the committed retained set against the independent runtime-under-test.
+pub fn verify() -> RetentionReport {
+    let (set, present) = load();
+    let regressions = set
+        .iter()
+        .filter(|f| !agrees(f, |inp| native_resolve(inp)))
+        .map(|f| f.master_seed)
+        .collect();
+    RetentionReport { present, total: set.len(), regressions }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use runtime_under_test::buggy_resolve;
+
+    #[test]
+    fn serialize_roundtrips() {
+        let set = vec![
+            RetainedFailure { master_seed: 0xABC, gate: "runtime_equivalence".into(), note: "dropped edge".into() },
+            RetainedFailure { master_seed: 0x1, gate: "replay".into(), note: "drift".into() },
+        ];
+        assert_eq!(parse(&serialize(&set)), set);
+    }
+
+    #[test]
+    fn committed_retained_set_exists_and_holds() {
+        // The committed file must be present (the retention mechanism is wired),
+        // and every retained counterexample must still be fixed.
+        let report = verify();
+        assert!(report.present, "committed retained_failures.tsv is missing");
+        assert!(report.regressions.is_empty(), "regressions: {:?}", report.regressions);
+    }
+
+    /// A retained counterexample re-checked against a *buggy* runtime must be
+    /// flagged as a regression. Proves the retention check is not vacuous: if a
+    /// fixed bug reappears, retention catches it.
+    #[test]
+    fn reintroduced_bug_is_caught_by_retention() {
+        let f = RetainedFailure { master_seed: 42, gate: "runtime_equivalence".into(), note: "synthetic".into() };
+        // Against the real runtime the case agrees (the bug is fixed)...
+        assert!(agrees(&f, |inp| native_resolve(inp)));
+        // ...but a runtime that reintroduces the bug is caught as a regression.
+        assert!(!agrees(&f, |inp| buggy_resolve(inp)));
+    }
+}

crates/runtime_under_test/Cargo.toml

@@ -10,5 +10,8 @@ rune_ir = { path = "../rune_ir" }
 trace_model = { path = "../trace_model" }
 reference_runtime = { path = "../reference_runtime" }
 
+[features]
+negative_controls = []
+
 [lib]
 path = "src/lib.rs"

crates/runtime_under_test/src/lib.rs

@@ -1,53 +1,52 @@
 //! `runtime_under_test` — the runtime that CI proves equivalent to the
-//! reference. It is configuration-driven: the canonical configuration must
+//! reference. Unlike the reference, this crate does **not** call the reference
-//! match the reference bit-for-bit, while semantic mutation swaps in a mutated
+//! engine: it carries its own independent interpreter ([`native::native_resolve`])
-//! configuration to verify the test suite can detect any divergence.
+//! re-derived from the spec. The runtime-equivalence gate therefore compares
-//!
+//! two genuinely separate implementations, so 100% agreement is *evidence* that
-//! Per the spec's mandatory order, the *optimized* runtime may not begin until
+//! the spec is implemented correctly rather than a tautology. A transcription
-//! steps 1–7 pass CI; until then this runtime is the reference engine driven
+//! error in either implementation surfaces as an equivalence failure (proven by
-//! through the same config surface, which is by construction equivalent.
+//! the negative-control test below).
 
-use reference_runtime::{execute, EngineConfig, ResolutionInput, ResolutionResult, Runtime};
+pub mod native;
 
-#[derive(Clone, Debug)]
+use reference_runtime::{ResolutionInput, ResolutionResult, Runtime};
-pub struct RuntimeUnderTest {
-    pub config: EngineConfig,
-}
 
-impl Default for RuntimeUnderTest {
+pub use native::native_resolve;
-    fn default() -> Self {
+
-        RuntimeUnderTest {
+#[derive(Clone, Debug, Default)]
-            config: EngineConfig::reference(),
+pub struct RuntimeUnderTest;
-        }
-    }
-}
 
 impl RuntimeUnderTest {
     pub fn new() -> Self {
-        Self::default()
+        RuntimeUnderTest
-    }
-
-    /// Construct with a specific engine config (used by semantic mutation to
-    /// install a mutated artifact).
-    pub fn with_config(config: EngineConfig) -> Self {
-        RuntimeUnderTest { config }
     }
 }
 
 impl Runtime for RuntimeUnderTest {
     fn resolve(&self, input: ResolutionInput) -> ResolutionResult {
-        execute(&self.config, &input)
+        native_resolve(&input)
     }
 }
 
+/// A deliberately broken runtime used as a negative control: it shares the
+/// independent interpreter but corrupts one recorded value. The equivalence
+/// gate **must** reject it. This proves the gate can fail.
+#[cfg(any(test, feature = "negative_controls"))]
+pub fn buggy_resolve(input: &ResolutionInput) -> ResolutionResult {
+    let mut r = native_resolve(input);
+    // Drop a single causal edge — a subtle bug an honest gate has to catch.
+    r.trace.causal_graph.edges.pop();
+    r
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
-    use reference_runtime::{canonical, execute, ReferenceRuntime};
+    use reference_runtime::{canonical, execute, EngineConfig};
-    use rune_ir::{Op, RuneProgram, RuneToken};
+    use rune_ir::{Op, RuneProgram, RuneToken, ALL_OPS};
     use world_model::{standard_executors, ProgramId, Rng, WorldId, WorldSnapshot, NUM_DOMAINS};
 
-    fn random_input(seed: u64) -> ResolutionInput {
+    fn rich_input(seed: u64) -> ResolutionInput {
         let mut rng = Rng::new(seed);
         let mut w = WorldSnapshot::blank(WorldId(seed), seed);
         for d in &mut w.domains {
@@ -63,9 +62,9 @@ mod tests {
                 w.causal_state.coupling[j][i] = rng.range_i64(-17, 17);
             }
         }
-        let tokens: Vec<RuneToken> = (0..30)
+        let tokens: Vec<RuneToken> = (0..40)
-            .map(|_| RuneToken {
+            .map(|i| RuneToken {
-                op: Op::from_u8(rng.next_u64() as u8),
+                op: if i % 3 == 0 { ALL_OPS[i % 12] } else { Op::from_u8(rng.next_u64() as u8) },
                 a: rng.next_u64() as u8,
                 b: rng.next_u64() as u8,
                 c: rng.next_u64() as u8,
@@ -75,25 +74,40 @@ mod tests {
         ResolutionInput {
             world: w,
             program: RuneProgram { id: ProgramId(seed), tokens, seed },
-            contexts: standard_executors(seed, 3),
+            contexts: standard_executors(seed, 4),
             contract_seed: seed,
             perturbation_seed: seed,
         }
     }
 
+    /// The core honesty property: the independent interpreter reproduces the
+    /// reference engine bit-for-bit over a large seed sweep. This is what makes
+    /// the equivalence gate meaningful rather than vacuous.
     #[test]
-    fn rut_matches_reference() {
+    fn native_matches_reference_bit_for_bit() {
-        use reference_runtime::Runtime;
+        let cfg = EngineConfig::reference();
-        let rut = RuntimeUnderTest::new();
+        for s in 0..2000u64 {
-        let reference = ReferenceRuntime::new();
+            let input = rich_input(s.wrapping_mul(0x9e3779b97f4a7c15) ^ 0xabc);
-        for s in 0..200 {
+            let a = canonical(&execute(&cfg, &input));
-            let input = random_input(s);
+            let b = canonical(&native_resolve(&input));
-            let a = canonical(&reference.resolve(input.clone()));
+            assert_eq!(a, b, "independent interpreter diverged at seed {s}");
-            let b = canonical(&rut.resolve(input.clone()));
-            assert_eq!(a, b, "divergence at seed {s}");
-            // and against the raw engine path
-            let c = canonical(&execute(&EngineConfig::reference(), &input));
-            assert_eq!(a, c);
         }
     }
+
+    /// Negative control: a runtime with a real bug is rejected by the canonical
+    /// comparison. Proves the equivalence gate is not vacuous.
+    #[test]
+    fn buggy_runtime_is_rejected() {
+        let cfg = EngineConfig::reference();
+        let mut caught = 0;
+        for s in 0..200u64 {
+            let input = rich_input(s + 1);
+            let a = canonical(&execute(&cfg, &input));
+            let b = canonical(&buggy_resolve(&input));
+            if a != b {
+                caught += 1;
+            }
+        }
+        assert!(caught > 0, "the equivalence gate failed to catch a buggy runtime");
+    }
 }

crates/runtime_under_test/src/native.rs

@@ -0,0 +1,503 @@
+//! An *independent* interpreter for the canonical engine behavior.
+//!
+//! This is the whole point of the runtime-equivalence gate: if the runtime
+//! under test merely called `reference_runtime::execute`, agreement would be a
+//! tautology and a bug in the shared interpreter would hide in both. This file
+//! re-derives the executable spec's canonical behavior from scratch, in a
+//! different code organization (a register-machine `Vm` rather than the
+//! reference's free-function dispatch), depending only on the shared *data*
+//! crates (`world_model`, `trace_model`, `rune_ir`) and never on the
+//! reference's engine. When the two implementations agree it is evidence; when
+//! a transcription error is introduced, the equivalence gate catches it (see
+//! the negative-control tests).
+//!
+//! Because the runtime under test only ever needs to reproduce the *canonical*
+//! reference configuration, the engine constants are inlined here as literals —
+//! they are the spec, independently restated, not imported.
+
+use reference_runtime::{ResolutionInput, ResolutionResult};
+use rune_ir::{Op, RuneProgram, RuneToken};
+use trace_model::{
+    BehaviorFingerprint, CausalEdge, CausalGraph, CausalNode, DivergenceGraph, DomainAccessGraph,
+    ExecutionTrace, FaultCode, FaultLog, InformationFlowGraph, PerturbationResponse, ReplayRecord,
+    TemporalGraph,
+};
+use world_model::{
+    DomainKind, ExecutionContext, Hash, Hasher, ScheduledEffect, WorldDelta, WorldSnapshot,
+    DomainId, HIDDEN_LANES, LANES, NUM_DOMAINS, REGS,
+};
+
+// --- The canonical engine constants, independently restated. ---------------
+const C1: u64 = 0xff51afd7ed558ccd;
+const C2: u64 = 0xc4ceb9fe1a85ec53;
+const S1: u32 = 33;
+const S2: u32 = 29;
+const S3: u32 = 32;
+const FUTURE_TURNS: usize = 3;
+const DIFFUSE_SPAN: usize = 3;
+
+#[inline]
+fn avalanche(z: i64) -> i64 {
+    let mut u = z as u64;
+    u ^= u >> S1;
+    u = u.wrapping_mul(C1);
+    u ^= u >> S2;
+    u = u.wrapping_mul(C2);
+    u ^= u >> S3;
+    u as i64
+}
+
+#[inline]
+fn combine(ctx: &ExecutionContext, a: i64, b: i64, coupling: i64, kc: u64) -> i64 {
+    let mut z = a.wrapping_mul(kc as i64);
+    z ^= b.rotate_left(((kc & 31) as u32) + 1);
+    z = z.wrapping_add(coupling.wrapping_mul(b & 0xffff));
+    z ^= ctx.salt() as i64;
+    z = z.wrapping_add(ctx.profile.bias);
+    z = z.rotate_left((ctx.profile.rotate % 63) + 1);
+    avalanche(z)
+}
+
+/// A register-machine view of one single-executor run. Holds the working world,
+/// the trace graphs accumulated as the program executes, and the accumulator
+/// file. Organised as a stateful object with `&mut self` methods, deliberately
+/// unlike the reference's stateless free functions.
+struct Vm<'a> {
+    w: WorldSnapshot,
+    ctx: &'a ExecutionContext,
+    read_graph: DomainAccessGraph,
+    write_graph: DomainAccessGraph,
+    causal_graph: CausalGraph,
+    info_flow: InformationFlowGraph,
+    temporal: TemporalGraph,
+    faults: FaultLog,
+    acc: [i64; REGS],
+    acc_src: [usize; REGS],
+}
+
+impl<'a> Vm<'a> {
+    fn new(world: &WorldSnapshot, ctx: &'a ExecutionContext) -> Self {
+        let acc = world.execution_state.accumulator;
+        Vm {
+            w: world.clone(),
+            ctx,
+            read_graph: DomainAccessGraph::default(),
+            write_graph: DomainAccessGraph::default(),
+            causal_graph: CausalGraph::default(),
+            info_flow: InformationFlowGraph::default(),
+            temporal: TemporalGraph::default(),
+            faults: FaultLog::default(),
+            acc,
+            acc_src: [0; REGS],
+        }
+    }
+
+    #[inline]
+    fn rd(&self, dom: usize, lane: usize, hidden: bool) -> i64 {
+        if hidden {
+            self.w.domains[dom].hidden[lane % HIDDEN_LANES]
+        } else {
+            self.w.domains[dom].observed[lane % LANES]
+        }
+    }
+
+    #[inline]
+    fn wr(&mut self, dom: usize, lane: usize, hidden: bool, val: i64) {
+        if hidden {
+            self.w.domains[dom].hidden[lane % HIDDEN_LANES] = val;
+        } else {
+            self.w.domains[dom].observed[lane % LANES] = val;
+        }
+    }
+
+    #[inline]
+    fn coupling(&self, to: usize, from: usize) -> i64 {
+        self.w.causal_state.coupling[to][from]
+    }
+
+    #[inline]
+    fn mix(&self, a: i64, b: i64, coupling: i64, kc: u64) -> i64 {
+        combine(self.ctx, a, b, coupling, kc)
+    }
+
+    /// Record one data-movement edge across every graph, in the canonical order.
+    #[allow(clippy::too_many_arguments)]
+    fn flow(
+        &mut self,
+        from_dom: usize,
+        from_lane: usize,
+        from_hidden: bool,
+        to_dom: usize,
+        to_lane: usize,
+        to_hidden: bool,
+        step: u32,
+        weight: i64,
+    ) {
+        self.read_graph.access_count[from_dom] += 1;
+        self.write_graph.access_count[to_dom] += 1;
+        self.read_graph.edges.push((from_dom as u8, to_dom as u8, 1));
+        self.write_graph.edges.push((from_dom as u8, to_dom as u8, 1));
+        self.info_flow
+            .edges
+            .push((from_dom as u8, to_dom as u8, (weight as u64).count_ones()));
+        self.causal_graph.edges.push(CausalEdge {
+            from: CausalNode {
+                domain: from_dom as u8,
+                lane: from_lane as u8,
+                hidden: from_hidden,
+                step,
+            },
+            to: CausalNode {
+                domain: to_dom as u8,
+                lane: to_lane as u8,
+                hidden: to_hidden,
+                step,
+            },
+            weight,
+        });
+    }
+
+    fn run(&mut self, program: &RuneProgram) {
+        for (i, tok) in program.tokens.iter().enumerate() {
+            let step = i as u32;
+            self.step(tok, step);
+            let dst = tok.dst_domain();
+            let lane = tok.lane();
+            let r = (step as usize) % REGS;
+            self.acc[r] = self.acc[r].wrapping_add(self.w.domains[dst].observed[lane]);
+        }
+        self.w.execution_state.accumulator = self.acc;
+    }
+
+    fn step(&mut self, tok: &RuneToken, step: u32) {
+        let src = tok.src_domain();
+        let dst = tok.dst_domain();
+        let lane = tok.lane();
+        let lane2 = tok.lane2();
+        let kc = DomainKind::from_index(dst).mix_const();
+        let coupling = self.coupling(dst, src);
+
+        match tok.op {
+            Op::Mix => {
+                let a = self.rd(src, lane, false);
+                let b = self.rd(dst, lane2, false);
+                let v = self.mix(a, b, coupling, kc);
+                self.wr(dst, lane, false, v);
+                self.flow(src, lane, false, dst, lane, false, step, v);
+                self.flow(dst, lane2, false, dst, lane, false, step, v);
+            }
+            Op::Channel => {
+                let a = self.rd(src, lane, false);
+                let v = self.mix(a, coupling, coupling, kc);
+                self.wr(dst, lane2, false, v);
+                self.flow(src, lane, false, dst, lane2, false, step, v);
+            }
+            Op::Branch => {
+                let probe = self.rd(src, lane, false);
+                let take_hot =
+                    probe.wrapping_add(self.ctx.profile.bias) > self.ctx.profile.branch_threshold;
+                if take_hot {
+                    let b = self.rd(dst, lane, false);
+                    let v = self.mix(probe, b, coupling, kc);
+                    self.wr(dst, lane, false, v);
+                    self.flow(src, lane, false, dst, lane, false, step, v);
+                } else {
+                    let b = self.rd(dst, lane2, false);
+                    let v = self.mix(b, probe, coupling, kc).wrapping_add(0x5bd1e9);
+                    self.wr(dst, lane2, false, v);
+                    self.flow(src, lane, false, dst, lane2, false, step, v);
+                    self.faults.push(FaultCode::UnreachableBranch, step, 0);
+                }
+            }
+            Op::Schedule => {
+                let a = self.rd(src, lane, false);
+                let b = self.rd(dst, lane, false);
+                let v = self.mix(a, b, coupling, kc);
+                let offset = 1 + (tok.imm.rem_euclid(3)) as u8;
+                let hidden = tok.mode() & 1 == 1;
+                self.w.time_state.pending.push(ScheduledEffect {
+                    turn_offset: offset,
+                    domain: DomainId(dst as u8),
+                    lane,
+                    hidden,
+                    value: v,
+                });
+                self.temporal.edges.push((step, offset, dst as u8));
+                self.flow(src, lane, false, dst, lane, hidden, step, v);
+            }
+            Op::Resonate => {
+                let a = self.rd(src, lane, false);
+                let b = self.rd(dst, lane, false);
+                let m = self.mix(a, b, coupling, kc);
+                let va = a.wrapping_add(m);
+                let vb = b ^ m;
+                self.wr(src, lane, false, va);
+                self.wr(dst, lane, false, vb);
+                self.flow(dst, lane, false, src, lane, false, step, va);
+                self.flow(src, lane, false, dst, lane, false, step, vb);
+            }
+            Op::Observe => {
+                let reg = tok.mode() % REGS;
+                let mut z: i64 = self.acc[reg];
+                let proj = self.w.observed_projection();
+                for k in 0..3 {
+                    let d = (src + k) % NUM_DOMAINS;
+                    let idx = d * LANES + (lane + k) % LANES;
+                    let cpl = self.coupling(dst, d);
+                    z = self.mix(z, proj[idx], cpl, kc);
+                    self.flow(d, (lane + k) % LANES, false, dst, lane, true, step, z);
+                }
+                self.acc[reg] = z;
+                self.acc_src[reg] = src;
+                self.wr(dst, tok.mode() % HIDDEN_LANES, true, z);
+            }
+            Op::Collapse => {
+                let reg = tok.mode() % REGS;
+                let a = self.acc[reg];
+                let b = self.rd(dst, lane, false);
+                if a == 0 {
+                    self.faults.push(FaultCode::EmptyAccumulator, step, reg as i64);
+                }
+                let v = self.mix(a, b, coupling, kc);
+                self.wr(dst, lane, false, v);
+                let asrc = self.acc_src[reg];
+                self.flow(asrc, 0, true, dst, lane, false, step, v);
+            }
+            Op::Invert => {
+                let b = self.rd(dst, lane, false);
+                let mut v = avalanche((!b).wrapping_add(tok.imm));
+                v ^= self.ctx.salt() as i64;
+                v = v.wrapping_add(self.ctx.profile.bias);
+                self.wr(dst, lane, false, v);
+                self.flow(dst, lane, false, dst, lane, false, step, v);
+            }
+            Op::Diffuse => {
+                let a = self.rd(src, lane, false);
+                for k in 0..DIFFUSE_SPAN {
+                    let d = (src + 1 + k) % NUM_DOMAINS;
+                    let tl = (lane + k) % LANES;
+                    let prev = self.rd(d, tl, false);
+                    let cpl = self.coupling(d, src);
+                    let kc2 = DomainKind::from_index(d).mix_const();
+                    let v = self.mix(a, prev, cpl, kc2);
+                    self.wr(d, tl, false, prev.wrapping_add(v));
+                    self.flow(src, lane, false, d, tl, false, step, v);
+                }
+            }
+            Op::Anchor => {
+                let bound = (tok.imm.unsigned_abs() % 1_000_000) as i64 + 1;
+                let b = self.rd(dst, lane, false);
+                let diag = self.coupling(dst, dst);
+                let mut mixed = b.wrapping_add(diag);
+                mixed = mixed
+                    .wrapping_add(self.ctx.profile.bias)
+                    .wrapping_add((self.ctx.salt() & 0xffff) as i64);
+                let clamped = mixed.clamp(-bound, bound);
+                if clamped != mixed {
+                    self.faults.push(FaultCode::Saturated, step, bound);
+                }
+                self.wr(dst, lane, false, clamped);
+                self.flow(dst, lane, false, dst, lane, false, step, clamped);
+            }
+            Op::Echoback => {
+                let h = self.rd(dst, tok.mode() % HIDDEN_LANES, true);
+                let b = self.rd(dst, lane, false);
+                let v = self.mix(h, b, coupling, kc);
+                self.wr(dst, lane, false, v);
+                self.flow(dst, tok.mode() % HIDDEN_LANES, true, dst, lane, false, step, v);
+            }
+            Op::Imprint => {
+                let b = self.rd(dst, lane, false);
+                let hl = tok.mode() % HIDDEN_LANES;
+                let prevh = self.rd(dst, hl, true);
+                let v = self.mix(b, prevh, coupling, kc);
+                self.wr(dst, hl, true, v);
+                self.flow(dst, lane, false, dst, hl, true, step, v);
+            }
+        }
+    }
+}
+
+/// Advance the world one turn: resolve due scheduled effects, then run coupling
+/// diffusion. Identical semantics to the reference's `step_world`, restated.
+fn step_world(w: &mut WorldSnapshot) {
+    let pending = std::mem::take(&mut w.time_state.pending);
+    let mut still = Vec::new();
+    for e in pending {
+        if e.turn_offset <= 1 {
+            let d = e.domain.0 as usize;
+            if e.hidden {
+                let l = e.lane % HIDDEN_LANES;
+                w.domains[d].hidden[l] = w.domains[d].hidden[l].wrapping_add(e.value);
+            } else {
+                let l = e.lane % LANES;
+                w.domains[d].observed[l] = w.domains[d].observed[l].wrapping_add(e.value);
+            }
+        } else {
+            still.push(ScheduledEffect {
+                turn_offset: e.turn_offset - 1,
+                ..e
+            });
+        }
+    }
+    w.time_state.pending = still;
+
+    let snap = w.domains.clone();
+    for j in 0..NUM_DOMAINS {
+        for lane in 0..LANES {
+            let mut z = w.domains[j].observed[lane];
+            for i in 0..NUM_DOMAINS {
+                let c = w.causal_state.coupling[j][i];
+                z = z.wrapping_add(c.wrapping_mul(snap[i].observed[lane] & 0xff));
+            }
+            w.domains[j].observed[lane] = avalanche(z);
+        }
+        for hl in 0..HIDDEN_LANES {
+            let base = w.domains[j].hidden[hl].wrapping_add(snap[j].observed[0]);
+            w.domains[j].hidden[hl] = avalanche(base);
+        }
+    }
+    w.turn = w.turn.wrapping_add(1);
+}
+
+fn future_hash(start: &WorldSnapshot) -> Hash {
+    let mut w = start.clone();
+    let mut h = Hasher::new();
+    h.write_tag("future-3");
+    for _ in 0..FUTURE_TURNS {
+        step_world(&mut w);
+        for v in w.ground_truth() {
+            h.write_i64(v);
+        }
+    }
+    h.finish()
+}
+
+fn compute_divergence(finals: &[WorldSnapshot]) -> DivergenceGraph {
+    let n = finals.len();
+    let mut pairwise = vec![0.0f64; n * n];
+    let total = (NUM_DOMAINS * LANES) as f64;
+    for i in 0..n {
+        for j in 0..n {
+            if i == j {
+                continue;
+            }
+            let mut diff = 0usize;
+            for d in 0..NUM_DOMAINS {
+                for l in 0..LANES {
+                    if finals[i].domains[d].observed[l] != finals[j].domains[d].observed[l] {
+                        diff += 1;
+                    }
+                }
+            }
+            pairwise[i * n + j] = diff as f64 / total;
+        }
+    }
+    DivergenceGraph {
+        executor_count: n,
+        pairwise,
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+fn behavior_fingerprint(
+    delta: &WorldDelta,
+    causal: &CausalGraph,
+    read_graph: &DomainAccessGraph,
+    write_graph: &DomainAccessGraph,
+    info_flow: &InformationFlowGraph,
+    temporal: &TemporalGraph,
+    divergence: &DivergenceGraph,
+    future: Hash,
+) -> BehaviorFingerprint {
+    let mut features: Vec<i64> = Vec::new();
+    for dd in &delta.domain_deltas {
+        let mut s = 0i64;
+        for &v in &dd.observed {
+            s = s.wrapping_add(v);
+        }
+        features.push(s);
+    }
+    for dd in &delta.domain_deltas {
+        let mut s = 0i64;
+        for &v in &dd.hidden {
+            s = s.wrapping_add(v);
+        }
+        features.push(s);
+    }
+    features.push(causal.causal_rank() as i64);
+    features.push(causal.edge_count() as i64);
+    features.push(read_graph.touched_count() as i64);
+    features.push(write_graph.touched_count() as i64);
+    features.push(info_flow.total_bits() as i64);
+    features.push(temporal.edge_count() as i64);
+    features.push((divergence.mean_divergence() * 1_000_000.0) as i64);
+    features.push(future.0 as i64);
+    BehaviorFingerprint::from_features(features)
+}
+
+/// The independent implementation of the canonical resolution.
+pub fn native_resolve(input: &ResolutionInput) -> ResolutionResult {
+    let contexts = if input.contexts.is_empty() {
+        world_model::standard_executors(input.world.seed, 3)
+    } else {
+        input.contexts.clone()
+    };
+
+    let world0 = input.world.clone();
+
+    let mut finals: Vec<WorldSnapshot> = Vec::with_capacity(contexts.len());
+    let mut primary: Option<Vm> = None;
+    for (idx, ctx) in contexts.iter().enumerate() {
+        let mut vm = Vm::new(&world0, ctx);
+        vm.run(&input.program);
+        finals.push(vm.w.clone());
+        if idx == 0 {
+            primary = Some(vm);
+        }
+    }
+    let vm = primary.expect("at least one executor");
+
+    let delta = WorldDelta::between(&world0, &vm.w);
+    let divergence = compute_divergence(&finals);
+    let fhash = future_hash(&vm.w);
+    let behavior = behavior_fingerprint(
+        &delta,
+        &vm.causal_graph,
+        &vm.read_graph,
+        &vm.write_graph,
+        &vm.info_flow,
+        &vm.temporal,
+        &divergence,
+        fhash,
+    );
+
+    let trace = ExecutionTrace {
+        read_graph: vm.read_graph,
+        write_graph: vm.write_graph,
+        causal_graph: vm.causal_graph,
+        information_flow: vm.info_flow,
+        executor_divergence: divergence,
+        temporal_graph: vm.temporal,
+        perturbation_response: PerturbationResponse::default(),
+        behavior_fingerprint: behavior,
+    };
+    let trace_hash = trace.canonical_hash();
+    let delta_hash = delta.hash();
+    let replay = ReplayRecord {
+        world_seed: input.world.seed,
+        program_seed: input.program.seed,
+        contract_seed: input.contract_seed,
+        perturbation_seed: input.perturbation_seed,
+        trace_hash,
+        delta_hash,
+        future_hash: fhash,
+    };
+
+    ResolutionResult {
+        delta,
+        trace,
+        faults: vm.faults,
+        replay,
+    }
+}

crates/semantic_mutation/src/lib.rs

@@ -1,8 +1,14 @@
 //! `semantic_mutation` — structurally generate mutated runtimes and prove the
-//! test suite kills every one. A mutant is an [`EngineConfig`] (the runtime
+//! test suite kills every one **by the named acceptance gate it targets**.
-//! artifact) with exactly one behavior-affecting knob changed. Every mutant
+//!
-//! must fail at least one named acceptance gate; a survivor means the tests are
+//! A mutant is an [`EngineConfig`] (the runtime artifact) with exactly one
-//! invalid and blocks merge.
+//! behavior-affecting knob changed. The spec requires that every mutant fail at
+//! least one *named* acceptance gate. An earlier version of this crate only
+//! checked that a mutant's canonical output *differed* from the reference — a
+//! weaker, wrong condition that a mutant could satisfy without tripping the gate
+//! it is supposed to expose. This version runs the actual named gate against
+//! each mutant and requires that specific gate to fail. A mutant that does not
+//! trip its named gate is a survivor and blocks merge.
 
 use reference_runtime::{canonical, execute, Canonical, EngineConfig, ResolutionInput};
 use world_model::NUM_DOMAINS;
@@ -10,6 +16,13 @@ use world_model::NUM_DOMAINS;
 /// A mutant runtime artifact.
 pub type RuntimeArtifact = EngineConfig;
 
+// --- Gate thresholds, mirrored from the CI gate definitions. ----------------
+const CAUSAL_EDGES_MIN: f64 = 24.0;
+const CAUSAL_RANK_P95_MIN: f64 = 6.0;
+const DOMAIN_APPEARS_MIN: f64 = 0.35;
+const DOMAIN_MUTATED_MIN: f64 = 0.20;
+const FUTURE_ALT_MIN: f64 = 0.50;
+
 /// The named acceptance gate a mutant is expected to fail.
 #[derive(Clone, Copy, PartialEq, Eq, Debug)]
 pub enum DetectionClass {
@@ -23,8 +36,8 @@ impl DetectionClass {
     pub fn name(self) -> &'static str {
         match self {
             DetectionClass::RuntimeEquivalence => "runtime_equivalence",
-            DetectionClass::CausalGate => "causal_gate",
+            DetectionClass::CausalGate => "causal_rank/trace",
-            DetectionClass::TemporalGate => "temporal_gate",
+            DetectionClass::TemporalGate => "metamorphic_response/temporal",
             DetectionClass::DomainParticipation => "domain_participation",
         }
     }
@@ -36,9 +49,6 @@ pub trait SemanticMutator {
     fn expected_detection_reason(&self) -> DetectionClass;
 }
 
-/// Each generated mutant is also a [`SemanticMutator`]: applying it to any base
-/// artifact reproduces its single-knob change, and it names the gate it must
-/// fail. This ties the structural generator to the spec's trait surface.
 impl SemanticMutator for Mutant {
     fn mutate(&self, _base: &RuntimeArtifact) -> RuntimeArtifact {
         self.config.clone()
@@ -58,7 +68,8 @@ pub struct Mutant {
 }
 
 /// Build the `i`-th mutant deterministically from the reference artifact.
-/// Every mutant differs from the reference in exactly one behavioral knob.
+/// Every mutant differs from the reference in exactly one behavioral knob, and
+/// is tagged with the named gate that change must trip.
 pub fn mutant_for(i: usize) -> Mutant {
     let base = EngineConfig::reference();
     let mut cfg = base.clone();
@@ -145,7 +156,6 @@ pub fn mutant_for(i: usize) -> Mutant {
         }
     };
 
-    // Safety net: guarantee the mutant is not accidentally identical.
     if cfg == base {
         cfg.use_hidden = !cfg.use_hidden;
     }
@@ -169,20 +179,153 @@ pub fn reference_canon(inputs: &[ResolutionInput]) -> Vec<Canonical> {
     inputs.iter().map(|inp| canonical(&execute(&cfg, inp))).collect()
 }
 
-/// Returns `Some(case_index)` of the first execution where the mutant diverges
+// --- Named-gate evaluators. -------------------------------------------------
-/// from the reference (i.e. the mutant is killed), or `None` if it survives.
+//
-pub fn kill_index(
+// Each evaluator computes, for a given engine config over the input corpus, the
-    mutant: &EngineConfig,
+// metric a named CI gate checks, and returns whether that gate FAILS. The
-    inputs: &[ResolutionInput],
+// reference config must pass all of them (asserted in tests); each mutant must
-    reference: &[Canonical],
+// fail the one it targets.
-) -> Option<usize> {
+
-    for (i, inp) in inputs.iter().enumerate() {
+fn percentile(mut v: Vec<f64>, p: f64) -> f64 {
-        let c = canonical(&execute(mutant, inp));
+    if v.is_empty() {
-        if c != reference[i] {
+        return 0.0;
-            return Some(i);
+    }
+    v.sort_by(|a, b| a.partial_cmp(b).unwrap());
+    let idx = (((v.len() - 1) as f64) * p).round() as usize;
+    v[idx.min(v.len() - 1)]
+}
+
+fn median(v: Vec<f64>) -> f64 {
+    if v.is_empty() {
+        return 0.0;
+    }
+    let mut s = v;
+    s.sort_by(|a, b| a.partial_cmp(b).unwrap());
+    s[s.len() / 2]
+}
+
+/// True if the causal/trace gate fails under `cfg`.
+fn causal_gate_fails(cfg: &EngineConfig, inputs: &[ResolutionInput]) -> bool {
+    let edges: Vec<f64> = inputs
+        .iter()
+        .map(|inp| execute(cfg, inp).trace.causal_edge_count() as f64)
+        .collect();
+    let ranks: Vec<f64> = inputs
+        .iter()
+        .map(|inp| execute(cfg, inp).trace.causal_rank() as f64)
+        .collect();
+    median(edges) < CAUSAL_EDGES_MIN || percentile(ranks, 0.05) < CAUSAL_RANK_P95_MIN
+}
+
+/// True if the domain-participation gate fails under `cfg`.
+fn domain_gate_fails(cfg: &EngineConfig, inputs: &[ResolutionInput]) -> bool {
+    let n = inputs.len().max(1) as f64;
+    let mut appears = [0u32; NUM_DOMAINS];
+    let mut mutated = [0u32; NUM_DOMAINS];
+    for inp in inputs {
+        let r = execute(cfg, inp);
+        for d in 0..NUM_DOMAINS {
+            if r.trace.read_graph.access_count[d] > 0 || r.trace.write_graph.access_count[d] > 0 {
+                appears[d] += 1;
+            }
+        }
+        for dd in &r.delta.domain_deltas {
+            if !dd.is_zero() {
+                mutated[dd.domain.0 as usize] += 1;
+            }
         }
     }
-    None
+    (0..NUM_DOMAINS).any(|d| {
+        (appears[d] as f64 / n) < DOMAIN_APPEARS_MIN
+            || (mutated[d] as f64 / n) < DOMAIN_MUTATED_MIN
+    })
+}
+
+/// Apply a fixed structural perturbation (bump domain 0, observed lane 0).
+fn perturbed(input: &ResolutionInput) -> ResolutionInput {
+    let mut p = input.clone();
+    p.world.domains[0].observed[0] = p.world.domains[0].observed[0].wrapping_add(101);
+    p.world.mark_perturbed();
+    p
+}
+
+/// True if the temporal gate fails under `cfg`. The temporal gate asserts the
+/// runtime carries genuine 3-turn future dynamics: it must (a) record temporal
+/// edges, (b) have a future sensitive to perturbation, and (c) reproduce the
+/// reference's 3-turn future. Any of these failing fails the gate.
+fn temporal_gate_fails(
+    cfg: &EngineConfig,
+    inputs: &[ResolutionInput],
+    reference: &[Canonical],
+) -> bool {
+    // (a) temporal edges present.
+    let tedges: Vec<f64> = inputs
+        .iter()
+        .map(|inp| execute(cfg, inp).trace.temporal_graph.edge_count() as f64)
+        .collect();
+    if median(tedges) < 1.0 {
+        return true;
+    }
+    // (b) future sensitive to perturbation.
+    let mut altered = 0usize;
+    for inp in inputs {
+        let base_future = execute(cfg, inp).replay.future_hash;
+        let pert_future = execute(cfg, &perturbed(inp)).replay.future_hash;
+        if base_future != pert_future {
+            altered += 1;
+        }
+    }
+    let alt_rate = altered as f64 / inputs.len().max(1) as f64;
+    if alt_rate < FUTURE_ALT_MIN {
+        return true;
+    }
+    // (c) future matches the reference's 3-turn future on every input.
+    for (inp, ref_c) in inputs.iter().zip(reference) {
+        if execute(cfg, inp).replay.future_hash != ref_c.future_hash {
+            return true;
+        }
+    }
+    false
+}
+
+/// True if the runtime-equivalence gate fails under `cfg` (i.e. the mutant
+/// diverges from the reference canonical view on at least one input).
+fn equivalence_gate_fails(
+    cfg: &EngineConfig,
+    inputs: &[ResolutionInput],
+    reference: &[Canonical],
+) -> bool {
+    inputs
+        .iter()
+        .zip(reference)
+        .any(|(inp, ref_c)| canonical(&execute(cfg, inp)) != *ref_c)
+}
+
+/// Evaluate whether a mutant is killed by its **named** gate. Returns `None` if
+/// killed (the named gate fails), or `Some(reason)` describing the survival.
+pub fn survival_reason(
+    mutant: &Mutant,
+    inputs: &[ResolutionInput],
+    reference: &[Canonical],
+) -> Option<String> {
+    let killed = match mutant.expected {
+        DetectionClass::RuntimeEquivalence => {
+            equivalence_gate_fails(&mutant.config, inputs, reference)
+        }
+        DetectionClass::CausalGate => causal_gate_fails(&mutant.config, inputs),
+        DetectionClass::TemporalGate => temporal_gate_fails(&mutant.config, inputs, reference),
+        DetectionClass::DomainParticipation => domain_gate_fails(&mutant.config, inputs),
+    };
+    if killed {
+        None
+    } else {
+        Some(format!(
+            "mutant {} ({}) did not fail its named gate {}",
+            mutant.id,
+            mutant.name,
+            mutant.expected.name()
+        ))
+    }
 }
 
 /// Result of running the full mutation suite.
@@ -199,17 +342,16 @@ impl MutationOutcome {
     }
 }
 
-/// Run all mutants against the input corpus.
+/// Run all mutants against the input corpus, killing each by its named gate.
 pub fn run_suite(count: usize, inputs: &[ResolutionInput]) -> MutationOutcome {
     let reference = reference_canon(inputs);
     let mutants = generate_mutants(count);
     let mut killed = 0;
     let mut survivors = Vec::new();
     for m in &mutants {
-        if kill_index(&m.config, inputs, &reference).is_some() {
+        match survival_reason(m, inputs, &reference) {
-            killed += 1;
+            None => killed += 1,
-        } else {
+            Some(reason) => survivors.push((m.id, reason)),
-            survivors.push((m.id, m.name.clone()));
         }
     }
     MutationOutcome {
@@ -223,13 +365,13 @@ pub fn run_suite(count: usize, inputs: &[ResolutionInput]) -> MutationOutcome {
 mod tests {
     use super::*;
     use rune_ir::{RuneProgram, RuneToken, ALL_OPS};
-    use world_model::{standard_executors, ProgramId, Rng, WorldId, WorldSnapshot, NUM_DOMAINS};
+    use world_model::{standard_executors, ProgramId, Rng, WorldId, WorldSnapshot, LANES, NUM_DOMAINS};
 
     fn rich_input(seed: u64) -> ResolutionInput {
         let mut rng = Rng::new(seed);
         let mut w = WorldSnapshot::blank(WorldId(seed), seed);
         for d in &mut w.domains {
-            for l in 0..world_model::LANES {
+            for l in 0..LANES {
                 d.observed[l] = rng.range_i64(-5000, 5000);
             }
             for l in 0..world_model::HIDDEN_LANES {
@@ -241,7 +383,6 @@ mod tests {
                 w.causal_state.coupling[j][i] = rng.range_i64(-17, 17);
             }
         }
-        // cover every op and every domain
         let tokens: Vec<RuneToken> = (0..40)
             .map(|i| RuneToken {
                 op: ALL_OPS[i % ALL_OPS.len()],
@@ -260,6 +401,35 @@ mod tests {
         }
     }
 
+    fn corpus() -> Vec<ResolutionInput> {
+        (0..16).map(|s| rich_input(s + 1)).collect()
+    }
+
+    #[test]
+    fn reference_passes_every_named_gate() {
+        let inputs = corpus();
+        let reference = reference_canon(&inputs);
+        let cfg = EngineConfig::reference();
+        assert!(!causal_gate_fails(&cfg, &inputs), "reference fails causal gate");
+        assert!(!domain_gate_fails(&cfg, &inputs), "reference fails domain gate");
+        assert!(
+            !temporal_gate_fails(&cfg, &inputs, &reference),
+            "reference fails temporal gate"
+        );
+        assert!(
+            !equivalence_gate_fails(&cfg, &inputs, &reference),
+            "reference fails equivalence gate"
+        );
+    }
+
+    #[test]
+    fn no_mutant_survives_its_named_gate() {
+        let inputs = corpus();
+        let outcome = run_suite(520, &inputs);
+        assert!(outcome.passed(), "survivors: {:?}", outcome.survivors);
+        assert_eq!(outcome.killed, outcome.total);
+    }
+
     #[test]
     fn every_mutant_differs_from_reference() {
         let base = EngineConfig::reference();
@@ -267,12 +437,4 @@ mod tests {
             assert_ne!(mutant_for(i).config, base, "mutant {i} equals reference");
         }
     }
-
-    #[test]
-    fn no_mutant_survives() {
-        let inputs: Vec<ResolutionInput> = (0..12).map(|s| rich_input(s + 1)).collect();
-        let outcome = run_suite(520, &inputs);
-        assert!(outcome.passed(), "survivors: {:?}", outcome.survivors);
-        assert_eq!(outcome.killed, outcome.total);
-    }
 }

crates/server/Cargo.toml

@@ -0,0 +1,15 @@
+[package]
+name = "server"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[[bin]]
+name = "magicka-server"
+path = "src/main.rs"
+
+[dependencies]
+protocol = { path = "../protocol" }
+game_runtime = { path = "../game_runtime" }
+world_model = { path = "../world_model" }
+web_client = { path = "../web_client" }

crates/server/src/http.rs

@@ -0,0 +1,113 @@
+//! Minimal HTTP/1.1 request parsing — only enough to tell a static GET from a
+//! WebSocket upgrade and to read the upgrade key. Tolerant and total: a
+//! malformed request yields `None`, never a panic.
+
+use std::collections::BTreeMap;
+use std::io::{self, BufRead};
+
+/// A parsed request head.
+pub struct Request {
+    pub method: String,
+    pub path: String,
+    pub headers: BTreeMap<String, String>,
+}
+
+impl Request {
+    pub fn header(&self, name: &str) -> Option<&str> {
+        self.headers.get(&name.to_ascii_lowercase()).map(|s| s.as_str())
+    }
+
+    /// True if this is a WebSocket upgrade request.
+    pub fn is_websocket_upgrade(&self) -> bool {
+        self.header("upgrade")
+            .map(|v| v.eq_ignore_ascii_case("websocket"))
+            .unwrap_or(false)
+            && self
+                .header("connection")
+                .map(|v| v.to_ascii_lowercase().contains("upgrade"))
+                .unwrap_or(false)
+    }
+
+    pub fn websocket_key(&self) -> Option<&str> {
+        self.header("sec-websocket-key")
+    }
+}
+
+/// Read and parse the request head from a buffered reader. Returns `Ok(None)`
+/// on a clean EOF before any bytes.
+pub fn read_request<R: BufRead>(r: &mut R) -> io::Result<Option<Request>> {
+    let mut line = String::new();
+    let n = r.read_line(&mut line)?;
+    if n == 0 {
+        return Ok(None);
+    }
+    let mut parts = line.trim_end().split_whitespace();
+    let method = match parts.next() {
+        Some(m) => m.to_string(),
+        None => return Ok(None),
+    };
+    let path = parts.next().unwrap_or("/").to_string();
+
+    let mut headers = BTreeMap::new();
+    loop {
+        let mut h = String::new();
+        let hn = r.read_line(&mut h)?;
+        if hn == 0 {
+            break;
+        }
+        let trimmed = h.trim_end();
+        if trimmed.is_empty() {
+            break;
+        }
+        if let Some((k, v)) = trimmed.split_once(':') {
+            headers.insert(k.trim().to_ascii_lowercase(), v.trim().to_string());
+        }
+        // Bound header count defensively.
+        if headers.len() > 100 {
+            break;
+        }
+    }
+    Ok(Some(Request { method, path, headers }))
+}
+
+/// Build the 101 Switching Protocols handshake response.
+pub fn handshake_response(accept: &str) -> String {
+    format!(
+        "HTTP/1.1 101 Switching Protocols\r\n\
+         Upgrade: websocket\r\n\
+         Connection: Upgrade\r\n\
+         Sec-WebSocket-Accept: {accept}\r\n\r\n"
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::io::BufReader;
+
+    #[test]
+    fn parses_websocket_upgrade() {
+        let raw = "GET /ws HTTP/1.1\r\nHost: x\r\nUpgrade: websocket\r\nConnection: Upgrade\r\nSec-WebSocket-Key: abc\r\n\r\n";
+        let mut r = BufReader::new(raw.as_bytes());
+        let req = read_request(&mut r).unwrap().unwrap();
+        assert_eq!(req.method, "GET");
+        assert_eq!(req.path, "/ws");
+        assert!(req.is_websocket_upgrade());
+        assert_eq!(req.websocket_key(), Some("abc"));
+    }
+
+    #[test]
+    fn parses_plain_get() {
+        let raw = "GET /app.js HTTP/1.1\r\nHost: x\r\n\r\n";
+        let mut r = BufReader::new(raw.as_bytes());
+        let req = read_request(&mut r).unwrap().unwrap();
+        assert!(!req.is_websocket_upgrade());
+        assert_eq!(req.path, "/app.js");
+    }
+
+    #[test]
+    fn empty_input_is_none() {
+        let mut r = BufReader::new("".as_bytes());
+        assert!(read_request(&mut r).unwrap().is_none());
+    }
+}

crates/server/src/lib.rs

@@ -0,0 +1,606 @@
+//! `server` — the authoritative match server (Phase B of `plan2.md`), built on
+//! `std::net` with no external crates.
+//!
+//! Responsibilities the server owns: match state, the turn timer, collecting
+//! submitted actions, driving resolution through [`game_runtime`], visibility
+//! filtering, replay recording, and disconnect handling. The browser is served
+//! the embedded client and then speaks the `protocol` over a WebSocket.
+//!
+//! Authority guarantees enforced here and covered by tests:
+//! * **No panic on bad input** — every client packet is decoded with the total
+//!   `protocol` decoder; a failure becomes a `ValidationReport`, never a crash.
+//! * **Late input rejected deterministically** — a `SubmitTurn` for any turn
+//!   other than the live one, or after the deadline, is rejected with a stable
+//!   reason.
+//! * **Disconnect cannot corrupt a match** — a dropped connection simply stops
+//!   submitting; that player's turns default to `Wait` and the match continues.
+//! * **Client cannot mutate hidden state** — only intent is accepted, and the
+//!   hidden ground truth is never serialized to a client.
+
+pub mod http;
+pub mod ws;
+
+use std::collections::BTreeMap;
+use std::io::{BufReader, Write};
+use std::net::{SocketAddr, TcpListener, TcpStream};
+use std::sync::mpsc::{self, Sender};
+use std::sync::{Arc, Mutex};
+use std::thread;
+use std::time::{Duration, Instant};
+
+use game_runtime::{duel_roster, solo_roster, Match};
+use protocol::{
+    Action, ClientMessage, MatchId, PlayerId, ReplayTurn, RuneDiagnostics, ServerMessage,
+};
+
+/// How long the timer thread sleeps between ticks.
+const TICK: Duration = Duration::from_millis(40);
+/// Replay turns per `ReplayChunk`.
+const REPLAY_CHUNK: usize = 16;
+
+/// An outbound item for a single connection's writer thread. Routing every
+/// write through one thread keeps frames from interleaving.
+enum Out {
+    Text(String),
+    Pong(Vec<u8>),
+    Close,
+}
+
+/// One live match plus its scheduling and connection state.
+struct Session {
+    m: Match,
+    turn_len: Duration,
+    deadline: Instant,
+    pending: BTreeMap<u32, Action>,
+    conns: BTreeMap<u32, Sender<Out>>,
+    /// Entity ids that are human-controlled (vs. a dummy).
+    human_slots: Vec<u32>,
+}
+
+impl Session {
+    fn deadline_ms(&self, now: Instant) -> u64 {
+        self.deadline.saturating_duration_since(now).as_millis() as u64
+    }
+
+    fn snapshot_msg(&self, player: u32) -> ServerMessage {
+        ServerMessage::MatchState {
+            match_id: self.m.id,
+            player_id: PlayerId(player),
+            turn: self.m.turn,
+            snapshot: self.m.visible_for(player),
+        }
+    }
+}
+
+/// The shared manager: all matches, behind one mutex.
+pub struct Manager {
+    sessions: BTreeMap<u64, Session>,
+    next_auto_id: u64,
+    turn_ms: u64,
+}
+
+/// Result of a successful join.
+struct JoinOk {
+    match_id: MatchId,
+    player_id: u32,
+    initial: ServerMessage,
+    turn_started: ServerMessage,
+}
+
+impl Manager {
+    fn new(turn_ms: u64) -> Manager {
+        Manager { sessions: BTreeMap::new(), next_auto_id: 1, turn_ms }
+    }
+
+    fn turn_len(&self) -> Duration {
+        Duration::from_millis(self.turn_ms)
+    }
+
+    /// Join (or create) a match. `requested = None` creates a fresh solo match
+    /// (player + dummy). `requested = Some(id)` joins an existing duel by id, or
+    /// creates that duel and takes the first human slot.
+    fn join(
+        &mut self,
+        name: &str,
+        requested: Option<MatchId>,
+        tx: Sender<Out>,
+    ) -> Result<JoinOk, String> {
+        let now = Instant::now();
+        let turn_len = self.turn_len();
+        let key = match requested {
+            Some(m) => m.0,
+            None => {
+                let id = self.next_auto_id;
+                self.next_auto_id += 1;
+                id
+            }
+        };
+
+        // Create the session if absent.
+        if !self.sessions.contains_key(&key) {
+            let (roster, human_slots) = if requested.is_some() {
+                (duel_roster(name, "opponent"), vec![1u32, 2])
+            } else {
+                (solo_roster(name), vec![1u32])
+            };
+            let m = Match::new(MatchId(key), key, roster);
+            self.sessions.insert(
+                key,
+                Session {
+                    m,
+                    turn_len,
+                    deadline: now + turn_len,
+                    pending: BTreeMap::new(),
+                    conns: BTreeMap::new(),
+                    human_slots,
+                },
+            );
+        }
+
+        let session = self.sessions.get_mut(&key).unwrap();
+        // Find the first human slot without a live connection.
+        let slot = session
+            .human_slots
+            .iter()
+            .copied()
+            .find(|s| !session.conns.contains_key(s))
+            .ok_or_else(|| "match is full".to_string())?;
+
+        // Adopt the player's chosen name on their entity.
+        if let Some(e) = session.m.entity_mut(slot) {
+            e.name = name.to_string();
+        }
+        session.conns.insert(slot, tx);
+
+        Ok(JoinOk {
+            match_id: MatchId(key),
+            player_id: slot,
+            initial: session.snapshot_msg(slot),
+            turn_started: ServerMessage::TurnStarted {
+                turn: session.m.turn,
+                deadline_ms: session.deadline_ms(now),
+            },
+        })
+    }
+
+    /// Queue a turn submission. Rejects late / wrong-turn submissions
+    /// deterministically.
+    fn submit(
+        &mut self,
+        match_key: u64,
+        player: u32,
+        turn: u64,
+        action: Action,
+    ) -> Result<(), String> {
+        let now = Instant::now();
+        let session = self
+            .sessions
+            .get_mut(&match_key)
+            .ok_or_else(|| "no such match".to_string())?;
+        if session.m.finished {
+            return Err("match has ended".to_string());
+        }
+        if turn != session.m.turn {
+            return Err(format!(
+                "wrong turn: submitted {}, live turn is {}",
+                turn, session.m.turn
+            ));
+        }
+        if now > session.deadline {
+            return Err("late: turn deadline has passed".to_string());
+        }
+        session.pending.insert(player, action);
+        Ok(())
+    }
+
+    fn set_program(
+        &mut self,
+        match_key: u64,
+        player: u32,
+        tokens: Vec<protocol::RuneTokenWire>,
+    ) -> Result<RuneDiagnostics, String> {
+        let session = self
+            .sessions
+            .get_mut(&match_key)
+            .ok_or_else(|| "no such match".to_string())?;
+        session.m.set_program(player, tokens);
+        Ok(session.m.diagnostics_for_player(player))
+    }
+
+    fn inspect(&self, match_key: u64, target: u32) -> Result<ServerMessage, String> {
+        let session = self
+            .sessions
+            .get(&match_key)
+            .ok_or_else(|| "no such match".to_string())?;
+        let diagnostics = session.m.diagnostics_for_player(target);
+        Ok(ServerMessage::ObservationResult { target, diagnostics })
+    }
+
+    fn replay_chunks(&self, match_key: u64) -> Result<Vec<ServerMessage>, String> {
+        let session = self
+            .sessions
+            .get(&match_key)
+            .ok_or_else(|| "no such match".to_string())?;
+        let m = &session.m;
+        let turns: Vec<ReplayTurn> = m
+            .replay
+            .turns
+            .iter()
+            .map(|rt| ReplayTurn {
+                turn: rt.turn,
+                inputs: rt
+                    .inputs
+                    .iter()
+                    .map(|i| (i.player, i.action.clone()))
+                    .collect(),
+                runtime_hash: format!("{}", rt.turn_hash),
+            })
+            .collect();
+        let final_hash = m.final_hash_hex();
+        let chunks: Vec<&[ReplayTurn]> = if turns.is_empty() {
+            vec![&[]]
+        } else {
+            turns.chunks(REPLAY_CHUNK).collect()
+        };
+        let total = chunks.len() as u32;
+        Ok(chunks
+            .into_iter()
+            .enumerate()
+            .map(|(i, c)| ServerMessage::ReplayChunk {
+                match_id: m.id,
+                seed: m.seed,
+                index: i as u32,
+                total,
+                turns: c.to_vec(),
+                final_hash: final_hash.clone(),
+            })
+            .collect())
+    }
+
+    /// Advance any match whose deadline has elapsed. Runs under the lock; sends
+    /// are non-blocking on unbounded channels.
+    fn tick(&mut self, now: Instant) {
+        let mut empty: Vec<u64> = Vec::new();
+        for (key, session) in self.sessions.iter_mut() {
+            if session.conns.is_empty() {
+                empty.push(*key);
+                continue;
+            }
+            if session.m.finished || now < session.deadline {
+                continue;
+            }
+            // Resolve the turn from queued submissions.
+            let subs: Vec<(u32, Action)> =
+                session.pending.iter().map(|(p, a)| (*p, a.clone())).collect();
+            let events = session.m.resolve_turn(&subs);
+            session.pending.clear();
+            let runtime_hash = session.m.last_turn_hash_hex();
+            let turn = session.m.turn;
+            // Broadcast the resolved state, filtered per player.
+            for (pid, tx) in session.conns.iter() {
+                let msg = ServerMessage::TurnResolved {
+                    turn,
+                    snapshot: session.m.visible_for(*pid),
+                    runtime_hash: runtime_hash.clone(),
+                    events: events.clone(),
+                };
+                let _ = tx.send(Out::Text(msg.encode()));
+            }
+            // Open the next turn unless the match just ended.
+            if !session.m.finished {
+                session.deadline = now + session.turn_len;
+                let ts = ServerMessage::TurnStarted {
+                    turn,
+                    deadline_ms: session.turn_len.as_millis() as u64,
+                };
+                for tx in session.conns.values() {
+                    let _ = tx.send(Out::Text(ts.encode()));
+                }
+            }
+        }
+        // Drop sessions nobody is connected to (replay no longer reachable).
+        for key in empty {
+            self.sessions.remove(&key);
+        }
+    }
+
+    fn disconnect(&mut self, match_key: u64, player: u32) {
+        if let Some(session) = self.sessions.get_mut(&match_key) {
+            session.conns.remove(&player);
+            if session.conns.is_empty() {
+                self.sessions.remove(&match_key);
+            }
+        }
+    }
+}
+
+/// Configuration.
+pub struct Config {
+    pub turn_ms: u64,
+}
+
+impl Config {
+    pub fn from_env() -> Config {
+        let turn_ms = std::env::var("MAGICKA_TURN_MS")
+            .ok()
+            .and_then(|v| v.parse().ok())
+            .unwrap_or(5000);
+        Config { turn_ms }
+    }
+}
+
+/// Start the server on `addr`. Returns the bound address (useful when binding
+/// to port 0 in tests). Spawns the accept loop and the turn-timer thread as
+/// detached background threads.
+pub fn serve(addr: &str, cfg: Config) -> std::io::Result<SocketAddr> {
+    let listener = TcpListener::bind(addr)?;
+    let local = listener.local_addr()?;
+    let manager = Arc::new(Mutex::new(Manager::new(cfg.turn_ms)));
+
+    // Turn timer.
+    {
+        let mgr = Arc::clone(&manager);
+        thread::spawn(move || loop {
+            thread::sleep(TICK);
+            let now = Instant::now();
+            lock(&mgr).tick(now);
+        });
+    }
+
+    // Accept loop.
+    {
+        let mgr = Arc::clone(&manager);
+        thread::spawn(move || {
+            for stream in listener.incoming() {
+                if let Ok(stream) = stream {
+                    let mgr = Arc::clone(&mgr);
+                    thread::spawn(move || {
+                        let _ = handle_conn(stream, mgr);
+                    });
+                }
+            }
+        });
+    }
+
+    Ok(local)
+}
+
+/// Blocking entry point for the binary.
+pub fn run(addr: &str) -> std::io::Result<()> {
+    let local = serve(addr, Config::from_env())?;
+    eprintln!("magicka-server listening on http://{local}  (open it in a browser)");
+    loop {
+        thread::sleep(Duration::from_secs(3600));
+    }
+}
+
+fn handle_conn(stream: TcpStream, mgr: Arc<Mutex<Manager>>) -> std::io::Result<()> {
+    stream.set_nodelay(true).ok();
+    let mut head_reader = BufReader::new(stream.try_clone()?);
+    let req = match http::read_request(&mut head_reader)? {
+        Some(r) => r,
+        None => return Ok(()),
+    };
+
+    if !req.is_websocket_upgrade() {
+        // Static asset.
+        let mut s = stream;
+        let resp = web_client::http_response(&req.path).unwrap_or_else(web_client::not_found);
+        s.write_all(&resp)?;
+        return Ok(());
+    }
+
+    // Complete the WebSocket handshake.
+    let key = match req.websocket_key() {
+        Some(k) => k,
+        None => return Ok(()),
+    };
+    let accept = ws::accept_key(key);
+    {
+        let mut s = stream.try_clone()?;
+        s.write_all(http::handshake_response(&accept).as_bytes())?;
+        s.flush()?;
+    }
+
+    // Writer thread: the only thing that ever writes to this socket.
+    let (tx, rx) = mpsc::channel::<Out>();
+    let mut write_stream = stream.try_clone()?;
+    let writer = thread::spawn(move || {
+        for out in rx {
+            let r = match out {
+                Out::Text(s) => ws::write_text(&mut write_stream, &s),
+                Out::Pong(p) => ws::write_pong(&mut write_stream, &p),
+                Out::Close => {
+                    let _ = ws::write_close(&mut write_stream);
+                    break;
+                }
+            };
+            if r.is_err() {
+                break;
+            }
+        }
+    });
+
+    // Reader loop.
+    let mut read_stream = stream;
+    let mut match_key: Option<u64> = None;
+    let mut player_id: Option<u32> = None;
+
+    loop {
+        match ws::read_message(&mut read_stream) {
+            Ok(Some(ws::Message::Text(raw))) => {
+                dispatch(&mgr, &tx, &raw, &mut match_key, &mut player_id);
+            }
+            Ok(Some(ws::Message::Ping(p))) => {
+                let _ = tx.send(Out::Pong(p));
+            }
+            Ok(Some(ws::Message::Pong)) => {}
+            Ok(Some(ws::Message::Close)) | Ok(None) => break,
+            Err(_) => break,
+        }
+    }
+
+    // Disconnect: deregister so the match continues without corruption.
+    if let (Some(k), Some(p)) = (match_key, player_id) {
+        lock(&mgr).disconnect(k, p);
+    }
+    let _ = tx.send(Out::Close);
+    drop(tx);
+    let _ = writer.join();
+    Ok(())
+}
+
+/// Dispatch one decoded client message. Never panics: a decode failure or any
+/// rejected operation becomes a `ValidationReport`/`ErrorEvent`.
+fn dispatch(
+    mgr: &Arc<Mutex<Manager>>,
+    tx: &Sender<Out>,
+    raw: &str,
+    match_key: &mut Option<u64>,
+    player_id: &mut Option<u32>,
+) {
+    let msg = match ClientMessage::decode(raw) {
+        Ok(m) => m,
+        Err(e) => {
+            send(tx, ServerMessage::ValidationReport {
+                accepted: false,
+                detail: format!("malformed packet: {e}"),
+                diagnostics: RuneDiagnostics::default(),
+            });
+            return;
+        }
+    };
+
+    match msg {
+        ClientMessage::JoinMatch { name, match_id } => {
+            if player_id.is_some() {
+                send(tx, err_event("already_joined", "this connection already joined a match"));
+                return;
+            }
+            let mut m = lock(mgr);
+            match m.join(&name, match_id, tx.clone()) {
+                Ok(ok) => {
+                    *match_key = Some(ok.match_id.0);
+                    *player_id = Some(ok.player_id);
+                    send(tx, ok.initial);
+                    send(tx, ok.turn_started);
+                }
+                Err(detail) => send(tx, err_event("join_failed", &detail)),
+            }
+        }
+        ClientMessage::SubmitTurn { turn, action } => {
+            let (Some(k), Some(p)) = (*match_key, *player_id) else {
+                send(tx, err_event("not_joined", "join a match first"));
+                return;
+            };
+            let res = lock(mgr).submit(k, p, turn, action);
+            match res {
+                Ok(()) => send(tx, ServerMessage::ValidationReport {
+                    accepted: true,
+                    detail: format!("action queued for turn {turn}"),
+                    diagnostics: RuneDiagnostics::default(),
+                }),
+                Err(detail) => send(tx, ServerMessage::ValidationReport {
+                    accepted: false,
+                    detail,
+                    diagnostics: RuneDiagnostics::default(),
+                }),
+            }
+        }
+        ClientMessage::EditRuneProgram { tokens } => {
+            let (Some(k), Some(p)) = (*match_key, *player_id) else {
+                send(tx, err_event("not_joined", "join a match first"));
+                return;
+            };
+            match lock(mgr).set_program(k, p, tokens) {
+                Ok(diagnostics) => send(tx, ServerMessage::ValidationReport {
+                    accepted: true,
+                    detail: "program updated".to_string(),
+                    diagnostics,
+                }),
+                Err(detail) => send(tx, err_event("edit_failed", &detail)),
+            }
+        }
+        ClientMessage::InspectTarget { target } => {
+            let Some(k) = *match_key else {
+                send(tx, err_event("not_joined", "join a match first"));
+                return;
+            };
+            match lock(mgr).inspect(k, target) {
+                Ok(m) => send(tx, m),
+                Err(detail) => send(tx, err_event("inspect_failed", &detail)),
+            }
+        }
+        ClientMessage::RequestReplay { match_id } => {
+            match lock(mgr).replay_chunks(match_id.0) {
+                Ok(chunks) => {
+                    for c in chunks {
+                        send(tx, c);
+                    }
+                }
+                Err(detail) => send(tx, err_event("replay_failed", &detail)),
+            }
+        }
+        ClientMessage::Ping { .. } => {
+            // Liveness only; the WebSocket layer already handles control pings.
+        }
+    }
+}
+
+fn send(tx: &Sender<Out>, msg: ServerMessage) {
+    let _ = tx.send(Out::Text(msg.encode()));
+}
+
+/// Acquire the manager lock, recovering a poisoned guard. A panic in any single
+/// connection or tick must not permanently brick the server for everyone else.
+fn lock(mgr: &Arc<Mutex<Manager>>) -> std::sync::MutexGuard<'_, Manager> {
+    mgr.lock().unwrap_or_else(|p| p.into_inner())
+}
+
+fn err_event(code: &str, detail: &str) -> ServerMessage {
+    ServerMessage::ErrorEvent { code: code.to_string(), detail: detail.to_string() }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn manager_join_and_resolve_is_authoritative() {
+        let mut m = Manager::new(10);
+        let (tx, _rx) = mpsc::channel();
+        let ok = m.join("dev", None, tx).unwrap();
+        assert_eq!(ok.player_id, 1);
+        let key = ok.match_id.0;
+        // Submit a cast for the live turn.
+        assert!(m.submit(key, 1, 0, Action::Cast).is_ok());
+        // Wrong turn is rejected deterministically.
+        let e = m.submit(key, 1, 99, Action::Cast).unwrap_err();
+        assert!(e.contains("wrong turn"), "{e}");
+    }
+
+    #[test]
+    fn disconnect_drops_session_when_last_leaves() {
+        let mut m = Manager::new(10);
+        let (tx, _rx) = mpsc::channel();
+        let ok = m.join("dev", None, tx).unwrap();
+        let key = ok.match_id.0;
+        assert!(m.sessions.contains_key(&key));
+        m.disconnect(key, 1);
+        assert!(!m.sessions.contains_key(&key));
+    }
+
+    #[test]
+    fn duel_assigns_two_human_slots() {
+        let mut m = Manager::new(10);
+        let (tx1, _r1) = mpsc::channel();
+        let (tx2, _r2) = mpsc::channel();
+        let a = m.join("a", Some(MatchId(42)), tx1).unwrap();
+        let b = m.join("b", Some(MatchId(42)), tx2).unwrap();
+        assert_eq!(a.player_id, 1);
+        assert_eq!(b.player_id, 2);
+        // Third join to a full duel is rejected.
+        let (tx3, _r3) = mpsc::channel();
+        assert!(m.join("c", Some(MatchId(42)), tx3).is_err());
+    }
+}

crates/server/src/main.rs

@@ -0,0 +1,8 @@
+//! The Magicka VM web server binary. Serves the embedded browser client and the
+//! authoritative WebSocket protocol. Bind address via `MAGICKA_ADDR`
+//! (default `127.0.0.1:8080`); turn length via `MAGICKA_TURN_MS`.
+
+fn main() -> std::io::Result<()> {
+    let addr = std::env::var("MAGICKA_ADDR").unwrap_or_else(|_| "127.0.0.1:8080".to_string());
+    server::run(&addr)
+}

crates/server/src/ws.rs

@@ -0,0 +1,342 @@
+//! Minimal WebSocket (RFC 6455) support over `std::net`, no external crates.
+//! Implements just what the game needs: the upgrade handshake (SHA1 + base64),
+//! masked client-frame reading with fragment reassembly, and unmasked
+//! server-frame writing. All reads are length-checked so a hostile frame
+//! returns an `Err`, never a panic or unbounded allocation.
+
+use std::io::{self, Read, Write};
+
+const WS_GUID: &str = "258EAFA5-E914-47DA-95CA-C5AB0DC85B11";
+/// Reject any single message larger than this (defensive bound).
+pub const MAX_MESSAGE: usize = 1 << 20; // 1 MiB
+
+// ---------------------------------------------------------------------------
+// SHA-1 (FIPS 180-1). Used only for the handshake accept key.
+// ---------------------------------------------------------------------------
+
+fn sha1(data: &[u8]) -> [u8; 20] {
+    let mut h: [u32; 5] = [0x67452301, 0xEFCDAB89, 0x98BADCFE, 0x10325476, 0xC3D2E1F0];
+    let ml = (data.len() as u64) * 8;
+    let mut msg = data.to_vec();
+    msg.push(0x80);
+    while msg.len() % 64 != 56 {
+        msg.push(0);
+    }
+    msg.extend_from_slice(&ml.to_be_bytes());
+
+    for chunk in msg.chunks_exact(64) {
+        let mut w = [0u32; 80];
+        for (i, wi) in w.iter_mut().enumerate().take(16) {
+            *wi = u32::from_be_bytes([
+                chunk[i * 4],
+                chunk[i * 4 + 1],
+                chunk[i * 4 + 2],
+                chunk[i * 4 + 3],
+            ]);
+        }
+        for i in 16..80 {
+            w[i] = (w[i - 3] ^ w[i - 8] ^ w[i - 14] ^ w[i - 16]).rotate_left(1);
+        }
+        let (mut a, mut b, mut c, mut d, mut e) = (h[0], h[1], h[2], h[3], h[4]);
+        for (i, &wi) in w.iter().enumerate() {
+            let (f, k) = match i {
+                0..=19 => ((b & c) | ((!b) & d), 0x5A827999u32),
+                20..=39 => (b ^ c ^ d, 0x6ED9EBA1),
+                40..=59 => ((b & c) | (b & d) | (c & d), 0x8F1BBCDC),
+                _ => (b ^ c ^ d, 0xCA62C1D6),
+            };
+            let tmp = a
+                .rotate_left(5)
+                .wrapping_add(f)
+                .wrapping_add(e)
+                .wrapping_add(k)
+                .wrapping_add(wi);
+            e = d;
+            d = c;
+            c = b.rotate_left(30);
+            b = a;
+            a = tmp;
+        }
+        h[0] = h[0].wrapping_add(a);
+        h[1] = h[1].wrapping_add(b);
+        h[2] = h[2].wrapping_add(c);
+        h[3] = h[3].wrapping_add(d);
+        h[4] = h[4].wrapping_add(e);
+    }
+
+    let mut out = [0u8; 20];
+    for (i, hi) in h.iter().enumerate() {
+        out[i * 4..i * 4 + 4].copy_from_slice(&hi.to_be_bytes());
+    }
+    out
+}
+
+// ---------------------------------------------------------------------------
+// base64 (standard alphabet).
+// ---------------------------------------------------------------------------
+
+fn base64(data: &[u8]) -> String {
+    const ALPHABET: &[u8; 64] =
+        b"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+    let mut out = String::new();
+    for chunk in data.chunks(3) {
+        let b = [
+            chunk[0],
+            *chunk.get(1).unwrap_or(&0),
+            *chunk.get(2).unwrap_or(&0),
+        ];
+        let n = ((b[0] as u32) << 16) | ((b[1] as u32) << 8) | (b[2] as u32);
+        out.push(ALPHABET[((n >> 18) & 63) as usize] as char);
+        out.push(ALPHABET[((n >> 12) & 63) as usize] as char);
+        if chunk.len() > 1 {
+            out.push(ALPHABET[((n >> 6) & 63) as usize] as char);
+        } else {
+            out.push('=');
+        }
+        if chunk.len() > 2 {
+            out.push(ALPHABET[(n & 63) as usize] as char);
+        } else {
+            out.push('=');
+        }
+    }
+    out
+}
+
+/// Compute the `Sec-WebSocket-Accept` value for a client key.
+pub fn accept_key(client_key: &str) -> String {
+    let mut concat = client_key.to_string();
+    concat.push_str(WS_GUID);
+    base64(&sha1(concat.as_bytes()))
+}
+
+// ---------------------------------------------------------------------------
+// Frames.
+// ---------------------------------------------------------------------------
+
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+pub enum Opcode {
+    Continuation,
+    Text,
+    Binary,
+    Close,
+    Ping,
+    Pong,
+}
+
+impl Opcode {
+    fn from_u8(v: u8) -> Option<Opcode> {
+        Some(match v {
+            0x0 => Opcode::Continuation,
+            0x1 => Opcode::Text,
+            0x2 => Opcode::Binary,
+            0x8 => Opcode::Close,
+            0x9 => Opcode::Ping,
+            0xA => Opcode::Pong,
+            _ => return None,
+        })
+    }
+}
+
+struct Frame {
+    fin: bool,
+    opcode: Opcode,
+    payload: Vec<u8>,
+}
+
+fn read_frame<R: Read>(r: &mut R) -> io::Result<Frame> {
+    let mut hdr = [0u8; 2];
+    r.read_exact(&mut hdr)?;
+    let fin = hdr[0] & 0x80 != 0;
+    let opcode = Opcode::from_u8(hdr[0] & 0x0f)
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "bad opcode"))?;
+    let masked = hdr[1] & 0x80 != 0;
+    let len7 = (hdr[1] & 0x7f) as usize;
+    let len = match len7 {
+        126 => {
+            let mut b = [0u8; 2];
+            r.read_exact(&mut b)?;
+            u16::from_be_bytes(b) as usize
+        }
+        127 => {
+            let mut b = [0u8; 8];
+            r.read_exact(&mut b)?;
+            u64::from_be_bytes(b) as usize
+        }
+        n => n,
+    };
+    if len > MAX_MESSAGE {
+        return Err(io::Error::new(io::ErrorKind::InvalidData, "frame too large"));
+    }
+    // Per RFC, client frames MUST be masked.
+    let mask = if masked {
+        let mut m = [0u8; 4];
+        r.read_exact(&mut m)?;
+        Some(m)
+    } else {
+        None
+    };
+    let mut payload = vec![0u8; len];
+    r.read_exact(&mut payload)?;
+    if let Some(m) = mask {
+        for (i, b) in payload.iter_mut().enumerate() {
+            *b ^= m[i % 4];
+        }
+    }
+    Ok(Frame { fin, opcode, payload })
+}
+
+/// A complete application message read from the socket. Control frames are
+/// surfaced rather than answered inline so that *all* socket writes can be
+/// funneled through a single writer (avoiding interleaved frames when a server
+/// is both broadcasting and answering pings).
+pub enum Message {
+    Text(String),
+    /// A ping with its payload; the caller must reply with a pong.
+    Ping(Vec<u8>),
+    /// A pong (informational).
+    Pong,
+    /// The peer requested close.
+    Close,
+}
+
+/// Read one full WebSocket message, reassembling fragments. Returns `Ok(None)`
+/// on a clean EOF. Reads only — never writes to the socket.
+pub fn read_message<R: Read>(stream: &mut R) -> io::Result<Option<Message>> {
+    let mut buf: Vec<u8> = Vec::new();
+    let mut msg_op: Option<Opcode> = None;
+    loop {
+        let frame = match read_frame(stream) {
+            Ok(f) => f,
+            Err(e) if e.kind() == io::ErrorKind::UnexpectedEof => return Ok(None),
+            Err(e) => return Err(e),
+        };
+        match frame.opcode {
+            Opcode::Close => return Ok(Some(Message::Close)),
+            Opcode::Ping => return Ok(Some(Message::Ping(frame.payload))),
+            Opcode::Pong => return Ok(Some(Message::Pong)),
+            Opcode::Text | Opcode::Binary => {
+                if msg_op.is_some() {
+                    return Err(io::Error::new(io::ErrorKind::InvalidData, "interleaved frame"));
+                }
+                msg_op = Some(frame.opcode);
+                buf.extend_from_slice(&frame.payload);
+            }
+            Opcode::Continuation => {
+                if msg_op.is_none() {
+                    return Err(io::Error::new(io::ErrorKind::InvalidData, "stray continuation"));
+                }
+                buf.extend_from_slice(&frame.payload);
+            }
+        }
+        if buf.len() > MAX_MESSAGE {
+            return Err(io::Error::new(io::ErrorKind::InvalidData, "message too large"));
+        }
+        if frame.fin {
+            // We only surface text to the application; binary is decoded lossily.
+            let s = String::from_utf8_lossy(&buf).into_owned();
+            return Ok(Some(Message::Text(s)));
+        }
+    }
+}
+
+fn write_frame<W: Write>(w: &mut W, opcode: Opcode, payload: &[u8]) -> io::Result<()> {
+    let op = match opcode {
+        Opcode::Continuation => 0x0,
+        Opcode::Text => 0x1,
+        Opcode::Binary => 0x2,
+        Opcode::Close => 0x8,
+        Opcode::Ping => 0x9,
+        Opcode::Pong => 0xA,
+    };
+    let mut frame = vec![0x80 | op];
+    let len = payload.len();
+    if len < 126 {
+        frame.push(len as u8);
+    } else if len < 65536 {
+        frame.push(126);
+        frame.extend_from_slice(&(len as u16).to_be_bytes());
+    } else {
+        frame.push(127);
+        frame.extend_from_slice(&(len as u64).to_be_bytes());
+    }
+    frame.extend_from_slice(payload);
+    w.write_all(&frame)?;
+    w.flush()
+}
+
+/// Send a text message (server frames are never masked).
+pub fn write_text<W: Write>(w: &mut W, text: &str) -> io::Result<()> {
+    write_frame(w, Opcode::Text, text.as_bytes())
+}
+
+/// Send a pong frame echoing a ping payload.
+pub fn write_pong<W: Write>(w: &mut W, payload: &[u8]) -> io::Result<()> {
+    write_frame(w, Opcode::Pong, payload)
+}
+
+/// Send a close frame.
+pub fn write_close<W: Write>(w: &mut W) -> io::Result<()> {
+    write_frame(w, Opcode::Close, &[])
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn rfc_example_accept_key() {
+        // The canonical example from RFC 6455 section 1.3.
+        assert_eq!(
+            accept_key("dGhlIHNhbXBsZSBub25jZQ=="),
+            "s3pPLMBiTxaQ9kYGzzhZRbK+xOo="
+        );
+    }
+
+    #[test]
+    fn sha1_known_vector() {
+        // "abc" -> a9993e364706816aba3e25717850c26c9cd0d89d
+        let d = sha1(b"abc");
+        let hex: String = d.iter().map(|b| format!("{b:02x}")).collect();
+        assert_eq!(hex, "a9993e364706816aba3e25717850c26c9cd0d89d");
+    }
+
+    #[test]
+    fn base64_roundtrip_lengths() {
+        assert_eq!(base64(b""), "");
+        assert_eq!(base64(b"f"), "Zg==");
+        assert_eq!(base64(b"fo"), "Zm8=");
+        assert_eq!(base64(b"foo"), "Zm9v");
+        assert_eq!(base64(b"foobar"), "Zm9vYmFy");
+    }
+
+    #[test]
+    fn masked_text_frame_roundtrips_through_reader() {
+        use std::io::Cursor;
+        // Build a masked client text frame for "hi".
+        let payload = b"hi";
+        let mask = [0x01, 0x02, 0x03, 0x04];
+        let mut frame = vec![0x81, 0x80 | payload.len() as u8];
+        frame.extend_from_slice(&mask);
+        for (i, &b) in payload.iter().enumerate() {
+            frame.push(b ^ mask[i % 4]);
+        }
+        // Cursor implements Read+Write (write goes nowhere useful but pong path
+        // is not exercised here).
+        let mut cur = Cursor::new(frame);
+        match read_message(&mut cur).unwrap() {
+            Some(Message::Text(s)) => assert_eq!(s, "hi"),
+            _ => panic!("expected text"),
+        }
+    }
+
+    #[test]
+    fn oversized_frame_is_rejected() {
+        use std::io::Cursor;
+        // Declares a 127-length (8-byte) payload of u64::MAX — must error, not OOM.
+        let mut frame = vec![0x81, 0x80 | 127];
+        frame.extend_from_slice(&u64::MAX.to_be_bytes());
+        frame.extend_from_slice(&[0, 0, 0, 0]); // partial mask
+        let mut cur = Cursor::new(frame);
+        assert!(read_message(&mut cur).is_err());
+    }
+}

crates/trace_model/src/lib.rs

@@ -404,12 +404,216 @@ impl ExecutionTrace {
             ],
         )
     }
+
+    /// Serialize the FULL trace — every graph edge, count, and weight — to a
+    /// single line of whitespace-separated integers. This is the full-trace
+    /// *evidence* (not a summary): [`deserialize`] reconstructs the trace and
+    /// [`canonical_hash`] over the result is bit-identical, so an independent
+    /// verifier can recompute the trace hash from the raw structure rather than
+    /// trusting a reported digest. `f64` divergence values are stored as raw
+    /// bits for exact round-trip; the behavior-fingerprint hash is NOT stored —
+    /// it is re-derived from the features on load, so a fabricated digest cannot
+    /// survive.
+    pub fn serialize(&self) -> String {
+        let mut t: Vec<String> = vec!["trace-v1".to_string()];
+        let push_access = |t: &mut Vec<String>, g: &DomainAccessGraph| {
+            for c in &g.access_count {
+                t.push(c.to_string());
+            }
+            t.push(g.edges.len().to_string());
+            for &(a, b, w) in &g.edges {
+                t.push(a.to_string());
+                t.push(b.to_string());
+                t.push(w.to_string());
+            }
+        };
+        push_access(&mut t, &self.read_graph);
+        push_access(&mut t, &self.write_graph);
+        // causal
+        t.push(self.causal_graph.edges.len().to_string());
+        for e in &self.causal_graph.edges {
+            for v in [
+                e.from.domain as i64, e.from.lane as i64, e.from.hidden as i64, e.from.step as i64,
+                e.to.domain as i64, e.to.lane as i64, e.to.hidden as i64, e.to.step as i64, e.weight,
+            ] {
+                t.push(v.to_string());
+            }
+        }
+        // information flow
+        t.push(self.information_flow.edges.len().to_string());
+        for &(a, b, w) in &self.information_flow.edges {
+            t.push(a.to_string());
+            t.push(b.to_string());
+            t.push(w.to_string());
+        }
+        // divergence (f64 as raw bits)
+        t.push(self.executor_divergence.executor_count.to_string());
+        t.push(self.executor_divergence.pairwise.len().to_string());
+        for &v in &self.executor_divergence.pairwise {
+            t.push(v.to_bits().to_string());
+        }
+        // temporal
+        t.push(self.temporal_graph.edges.len().to_string());
+        for &(s, off, d) in &self.temporal_graph.edges {
+            t.push(s.to_string());
+            t.push(off.to_string());
+            t.push(d.to_string());
+        }
+        // perturbation response
+        for v in [
+            self.perturbation_response.total,
+            self.perturbation_response.altered_trace,
+            self.perturbation_response.altered_delta,
+            self.perturbation_response.altered_future,
+            self.perturbation_response.neutral_unexplained,
+        ] {
+            t.push(v.to_string());
+        }
+        // behavior features (fingerprint hash re-derived on load)
+        t.push(self.behavior_fingerprint.features.len().to_string());
+        for &f in &self.behavior_fingerprint.features {
+            t.push(f.to_string());
+        }
+        t.join(" ")
+    }
+
+    /// Reconstruct a trace from [`serialize`]. Total: returns `None` on any
+    /// malformed input rather than panicking.
+    pub fn deserialize(s: &str) -> Option<ExecutionTrace> {
+        let mut it = s.split_whitespace();
+        if it.next()? != "trace-v1" {
+            return None;
+        }
+        let nu = |it: &mut std::str::SplitWhitespace| -> Option<u64> { it.next()?.parse().ok() };
+        let ni = |it: &mut std::str::SplitWhitespace| -> Option<i64> { it.next()?.parse().ok() };
+        let read_access = |it: &mut std::str::SplitWhitespace| -> Option<DomainAccessGraph> {
+            let mut access_count = [0u32; NUM_DOMAINS];
+            for c in access_count.iter_mut() {
+                *c = nu(it)? as u32;
+            }
+            let n = nu(it)? as usize;
+            let mut edges = Vec::with_capacity(n);
+            for _ in 0..n {
+                edges.push((nu(it)? as u8, nu(it)? as u8, nu(it)? as u32));
+            }
+            Some(DomainAccessGraph { access_count, edges })
+        };
+        let read_graph = read_access(&mut it)?;
+        let write_graph = read_access(&mut it)?;
+        // causal
+        let cn = nu(&mut it)? as usize;
+        let mut cedges = Vec::with_capacity(cn);
+        for _ in 0..cn {
+            let from = CausalNode {
+                domain: ni(&mut it)? as u8,
+                lane: ni(&mut it)? as u8,
+                hidden: ni(&mut it)? != 0,
+                step: ni(&mut it)? as u32,
+            };
+            let to = CausalNode {
+                domain: ni(&mut it)? as u8,
+                lane: ni(&mut it)? as u8,
+                hidden: ni(&mut it)? != 0,
+                step: ni(&mut it)? as u32,
+            };
+            let weight = ni(&mut it)?;
+            cedges.push(CausalEdge { from, to, weight });
+        }
+        // info flow
+        let fin = nu(&mut it)? as usize;
+        let mut fedges = Vec::with_capacity(fin);
+        for _ in 0..fin {
+            fedges.push((nu(&mut it)? as u8, nu(&mut it)? as u8, nu(&mut it)? as u32));
+        }
+        // divergence
+        let executor_count = nu(&mut it)? as usize;
+        let pn = nu(&mut it)? as usize;
+        let mut pairwise = Vec::with_capacity(pn);
+        for _ in 0..pn {
+            pairwise.push(f64::from_bits(nu(&mut it)?));
+        }
+        // temporal
+        let tn = nu(&mut it)? as usize;
+        let mut tedges = Vec::with_capacity(tn);
+        for _ in 0..tn {
+            tedges.push((nu(&mut it)? as u32, nu(&mut it)? as u8, nu(&mut it)? as u8));
+        }
+        // perturbation response
+        let pr = PerturbationResponse {
+            total: nu(&mut it)? as usize,
+            altered_trace: nu(&mut it)? as usize,
+            altered_delta: nu(&mut it)? as usize,
+            altered_future: nu(&mut it)? as usize,
+            neutral_unexplained: nu(&mut it)? as usize,
+        };
+        // behavior features
+        let bn = nu(&mut it)? as usize;
+        let mut features = Vec::with_capacity(bn);
+        for _ in 0..bn {
+            features.push(ni(&mut it)?);
+        }
+        Some(ExecutionTrace {
+            read_graph,
+            write_graph,
+            causal_graph: CausalGraph { edges: cedges },
+            information_flow: InformationFlowGraph { edges: fedges },
+            executor_divergence: DivergenceGraph { executor_count, pairwise },
+            temporal_graph: TemporalGraph { edges: tedges },
+            perturbation_response: pr,
+            // Re-derive the fingerprint hash from features (not from a stored digest).
+            behavior_fingerprint: BehaviorFingerprint::from_features(features),
+        })
+    }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
 
+    fn sample_trace() -> ExecutionTrace {
+        let mut read_graph = DomainAccessGraph::default();
+        read_graph.access_count[0] = 3;
+        read_graph.access_count[2] = 1;
+        read_graph.edges.push((0, 2, 5));
+        let mut write_graph = DomainAccessGraph::default();
+        write_graph.access_count[2] = 4;
+        write_graph.edges.push((0, 2, 7));
+        let causal_graph = CausalGraph {
+            edges: vec![CausalEdge {
+                from: CausalNode { domain: 0, lane: 1, hidden: false, step: 2 },
+                to: CausalNode { domain: 2, lane: 0, hidden: true, step: 2 },
+                weight: -1234,
+            }],
+        };
+        ExecutionTrace {
+            read_graph,
+            write_graph,
+            causal_graph,
+            information_flow: InformationFlowGraph { edges: vec![(0, 2, 9), (2, 3, 4)] },
+            executor_divergence: DivergenceGraph { executor_count: 3, pairwise: vec![0.0, 0.5, 0.25, 0.5, 0.0, 0.125, 0.25, 0.125, 0.0] },
+            temporal_graph: TemporalGraph { edges: vec![(1, 2, 3)] },
+            perturbation_response: PerturbationResponse::default(),
+            behavior_fingerprint: BehaviorFingerprint::from_features(vec![1, -2, 3, -4]),
+        }
+    }
+
+    #[test]
+    fn full_trace_serialize_roundtrips_canonical_hash() {
+        let t = sample_trace();
+        let s = t.serialize();
+        let back = ExecutionTrace::deserialize(&s).expect("deserialize");
+        // The reconstructed trace is structurally equal and hashes identically.
+        assert_eq!(t, back);
+        assert_eq!(t.canonical_hash(), back.canonical_hash());
+    }
+
+    #[test]
+    fn deserialize_is_total_on_garbage() {
+        for s in ["", "nope", "trace-v1 1 2", "trace-v1 x y z"] {
+            let _ = ExecutionTrace::deserialize(s); // must not panic
+        }
+    }
+
     #[test]
     fn rank_of_identity_is_full() {
         let id: Vec<Vec<f64>> = (0..5)

crates/web_assets/Cargo.toml

@@ -0,0 +1,5 @@
+[package]
+name = "web_assets"
+version.workspace = true
+edition.workspace = true
+license.workspace = true

crates/web_assets/assets/app.js

@@ -0,0 +1,337 @@
+"use strict";
+// Browser client for Magicka VM. The browser only ever sends INTENT; the server
+// is the sole authority. This file mirrors the `protocol` crate's wire format
+// (version 1, envelope {v,type,body}). It never simulates the world — it renders
+// exactly what the server says is observable.
+
+const PROTOCOL_VERSION = 1;
+const OPS = ["mix","channel","branch","schedule","resonate","observe",
+             "collapse","invert","diffuse","anchor","echoback","imprint"];
+
+const state = {
+  ws: null,
+  playerId: null,
+  matchId: null,
+  turn: 0,
+  deadline: 0,
+  locked: false,
+  snapshot: null,
+  selectedTarget: null,
+  program: [],
+  slots: [[], [], []],
+  activeSlot: 0,
+  liveHashes: {},      // turn -> runtime_hash seen live
+  replay: null,        // { turns: [...], final_hash, cursor }
+};
+
+// ---- wire helpers -------------------------------------------------------
+function send(type, body) {
+  if (!state.ws || state.ws.readyState !== WebSocket.OPEN) return;
+  state.ws.send(JSON.stringify({ v: PROTOCOL_VERSION, type, body }));
+}
+
+function connect() {
+  const proto = location.protocol === "https:" ? "wss" : "ws";
+  const ws = new WebSocket(`${proto}://${location.host}/ws`);
+  state.ws = ws;
+  ws.onopen = () => {
+    setConn(true);
+    send("JoinMatch", { name: "dev-" + Math.floor(Math.random() * 1000), match_id: null });
+  };
+  ws.onclose = () => { setConn(false); setTimeout(connect, 1000); };
+  ws.onerror = () => ws.close();
+  ws.onmessage = (ev) => {
+    let msg;
+    try { msg = JSON.parse(ev.data); } catch (_) { return; }
+    if (!msg || msg.v !== PROTOCOL_VERSION) return;
+    handle(msg.type, msg.body || {});
+  };
+}
+
+function setConn(on) {
+  const el = document.getElementById("conn");
+  el.textContent = on ? "connected" : "disconnected";
+  el.className = "badge " + (on ? "on" : "off");
+}
+
+// ---- server message handling -------------------------------------------
+function handle(type, body) {
+  switch (type) {
+    case "MatchState":
+      state.playerId = body.player_id;
+      state.matchId = body.match_id;
+      state.turn = body.turn;
+      state.snapshot = body.snapshot;
+      render();
+      break;
+    case "TurnStarted":
+      state.turn = body.turn;
+      state.deadline = Date.now() + (body.deadline_ms || 0);
+      state.locked = false;
+      renderTimer();
+      break;
+    case "TurnResolved":
+      state.turn = body.turn;
+      state.snapshot = body.snapshot;
+      state.liveHashes[body.turn] = body.runtime_hash;
+      (body.events || []).forEach((e) => log(`t${body.turn}: ${e}`));
+      showHashes();
+      render();
+      break;
+    case "ObservationResult":
+      renderDiagnostics(body.diagnostics, `target ${body.target}`);
+      break;
+    case "ValidationReport":
+      log((body.accepted ? "✓ " : "✗ ") + body.detail);
+      if (body.diagnostics) renderDiagnostics(body.diagnostics, "program");
+      break;
+    case "ReplayChunk":
+      loadReplay(body);
+      break;
+    case "ErrorEvent":
+      log(`! ${body.code}: ${body.detail}`);
+      break;
+  }
+}
+
+// ---- rendering ----------------------------------------------------------
+function render() {
+  if (!state.snapshot) return;
+  renderArena();
+  renderDomains();
+  renderLogHistory();
+  renderTimer();
+}
+
+function renderArena() {
+  const s = state.snapshot;
+  const arena = document.getElementById("arena");
+  arena.style.gridTemplateColumns = `repeat(${s.arena_w}, 34px)`;
+  arena.innerHTML = "";
+  const at = {};
+  (s.observed_entities || []).forEach((e) => { at[`${e.x},${e.y}`] = e; });
+  for (let y = 0; y < s.arena_h; y++) {
+    for (let x = 0; x < s.arena_w; x++) {
+      const cell = document.createElement("div");
+      cell.className = "cell";
+      const e = at[`${x},${y}`];
+      if (e) {
+        cell.textContent = e.is_self ? "@" : (e.is_dummy ? "▣" : "&");
+        if (e.is_self) cell.classList.add("self");
+        if (state.selectedTarget === e.id) cell.classList.add("target");
+        if (!e.alive) cell.classList.add("dead");
+        const hp = document.createElement("span");
+        hp.className = "hp"; hp.textContent = e.hp;
+        cell.appendChild(hp);
+        cell.title = `${e.name} (#${e.id}) hp ${e.hp}`;
+        cell.onclick = () => { state.selectedTarget = e.id; renderArena(); };
+      }
+      arena.appendChild(cell);
+    }
+  }
+}
+
+function renderDomains() {
+  const s = state.snapshot;
+  const wrap = document.getElementById("domains");
+  wrap.innerHTML = "";
+  (s.observed_domains || []).forEach((d) => {
+    const el = document.createElement("div");
+    el.className = "domain";
+    const name = document.createElement("div");
+    name.className = "name"; name.textContent = `${d.index}: ${d.name}`;
+    el.appendChild(name);
+    d.observed.forEach((v, i) => {
+      const lane = document.createElement("span");
+      const k = (d.knowledge && d.knowledge[i]) || "unknown";
+      lane.className = "lane " + k;
+      lane.textContent = v === null ? "▒" : v;
+      lane.title = k;
+      el.appendChild(lane);
+    });
+    wrap.appendChild(el);
+  });
+  document.getElementById("redactions").textContent =
+    `${s.hidden_state_redactions} hidden state values withheld (hidden lanes + masked observations)`;
+  const inf = document.getElementById("inferred");
+  inf.innerHTML = "";
+  (s.inferred_markers || []).forEach((m) => {
+    const li = document.createElement("li"); li.textContent = m; inf.appendChild(li);
+  });
+}
+
+function renderLogHistory() {
+  // History from the snapshot is authoritative; live events are appended too.
+  const known = state.snapshot.known_history || [];
+  const log = document.getElementById("log");
+  if (log.dataset.lastTurn !== String(state.turn)) {
+    log.dataset.lastTurn = String(state.turn);
+  }
+}
+
+function renderTimer() {
+  const t = document.getElementById("timer");
+  const remain = Math.max(0, Math.ceil((state.deadline - Date.now()) / 1000));
+  t.textContent = `turn ${state.turn} — ${remain}s ${state.locked ? "(locked)" : ""}`;
+}
+setInterval(() => {
+  if (state.deadline) {
+    if (Date.now() > state.deadline) state.locked = true;
+    renderTimer();
+  }
+}, 250);
+
+function renderDiagnostics(d, label) {
+  const body = document.getElementById("diag-body");
+  body.innerHTML = "";
+  const row = (k, v) => {
+    const div = document.createElement("div");
+    div.className = "diag-row";
+    div.innerHTML = `<span class="diag-key">${k}:</span> ${v}`;
+    body.appendChild(div);
+  };
+  row("for", label);
+  row("known reads", (d.known_reads || []).join(", ") || "—");
+  row("known writes", (d.known_writes || []).join(", ") || "—");
+  row("observed risks", (d.observed_risks || []).join(", ") || "none observed");
+  row("unknown listeners", `<span class="warn">${d.unknown_listeners || 0}</span> (writes you cannot observe)`);
+  row("previous outcomes", (d.previous_outcomes || []).join(" | ") || "—");
+}
+
+// ---- rune editor --------------------------------------------------------
+function renderTokens() {
+  const wrap = document.getElementById("tokens");
+  wrap.innerHTML = "";
+  state.program.forEach((t, i) => {
+    const el = document.createElement("span");
+    el.className = "token";
+    el.textContent = `${OPS[t.op % OPS.length]} ${t.a},${t.b},${t.c}#${t.imm}`;
+    el.title = "click to remove";
+    el.onclick = () => { state.program.splice(i, 1); renderTokens(); };
+    wrap.appendChild(el);
+  });
+}
+
+function renderLibrary() {
+  const wrap = document.getElementById("library");
+  wrap.innerHTML = "";
+  state.slots.forEach((slot, i) => {
+    const el = document.createElement("div");
+    el.className = "slot" + (i === state.activeSlot ? " active" : "");
+    el.textContent = `slot ${i + 1} (${slot.length})`;
+    el.onclick = () => {
+      state.slots[state.activeSlot] = state.program.slice();
+      state.activeSlot = i;
+      state.program = state.slots[i].slice();
+      renderTokens(); renderLibrary();
+    };
+    wrap.appendChild(el);
+  });
+}
+
+function initEditor() {
+  const sel = document.getElementById("op-select");
+  OPS.forEach((op, i) => {
+    const o = document.createElement("option");
+    o.value = i; o.textContent = op; sel.appendChild(o);
+  });
+  document.getElementById("btn-add").onclick = () => {
+    state.program.push({
+      op: parseInt(sel.value, 10),
+      a: clampByte("tok-a"), b: clampByte("tok-b"), c: clampByte("tok-c"),
+      imm: parseInt(document.getElementById("tok-imm").value, 10) || 0,
+    });
+    renderTokens();
+  };
+  document.getElementById("btn-clear").onclick = () => { state.program = []; renderTokens(); };
+  document.getElementById("btn-save").onclick = () => {
+    send("EditRuneProgram", { tokens: state.program });
+    log("saved program (" + state.program.length + " runes)");
+  };
+  renderTokens(); renderLibrary();
+}
+
+function clampByte(id) {
+  let v = parseInt(document.getElementById(id).value, 10) || 0;
+  return Math.max(0, Math.min(255, v));
+}
+
+// ---- actions ------------------------------------------------------------
+function submit(action) {
+  if (state.locked) { log("turn locked — submission rejected client-side"); return; }
+  send("SubmitTurn", { turn: state.turn, action });
+}
+
+function initActions() {
+  document.querySelectorAll("[data-move]").forEach((b) => {
+    b.onclick = () => {
+      const [dx, dy] = b.dataset.move.split(",").map((n) => parseInt(n, 10));
+      submit({ kind: "move", dx, dy });
+    };
+  });
+  document.getElementById("btn-cast").onclick = () => submit({ kind: "cast" });
+  document.getElementById("btn-wait").onclick = () => submit({ kind: "wait" });
+  document.getElementById("btn-attack").onclick = () => {
+    if (state.selectedTarget === null) { log("select a target first"); return; }
+    submit({ kind: "attack", target: state.selectedTarget });
+  };
+  document.getElementById("btn-inspect").onclick = () => {
+    if (state.selectedTarget === null) { log("select a target first"); return; }
+    send("InspectTarget", { target: state.selectedTarget });
+  };
+}
+
+// ---- replay -------------------------------------------------------------
+function initReplay() {
+  document.getElementById("btn-replay").onclick = () => {
+    if (state.matchId === null) return;
+    state.replay = null;
+    send("RequestReplay", { match_id: state.matchId });
+  };
+  document.getElementById("btn-replay-step").onclick = stepReplay;
+}
+
+function loadReplay(chunk) {
+  if (!state.replay) state.replay = { turns: [], final_hash: chunk.final_hash, cursor: 0, seed: chunk.seed };
+  state.replay.turns = state.replay.turns.concat(chunk.turns || []);
+  state.replay.final_hash = chunk.final_hash;
+  document.getElementById("replay-status").textContent =
+    `replay loaded: ${state.replay.turns.length} turns (seed ${chunk.seed})`;
+}
+
+function stepReplay() {
+  if (!state.replay || state.replay.cursor >= state.replay.turns.length) {
+    document.getElementById("replay-status").textContent = "replay complete";
+    return;
+  }
+  const rt = state.replay.turns[state.replay.cursor++];
+  // Verify browser replay event order/hash matches what we saw live (Phase G).
+  const live = state.liveHashes[rt.turn];
+  const ok = live === undefined || live === rt.runtime_hash;
+  log(`replay t${rt.turn}: hash ${rt.runtime_hash} ${ok ? "✓ matches live" : "✗ MISMATCH"}`);
+  const hd = document.getElementById("hashes");
+  hd.innerHTML += `<div class="${ok ? "hash-ok" : "hash-bad"}">t${rt.turn} ${rt.runtime_hash}</div>`;
+  document.getElementById("replay-status").textContent =
+    `replay turn ${rt.turn} / ${state.replay.turns.length}`;
+}
+
+function showHashes() {
+  const hd = document.getElementById("hashes");
+  hd.innerHTML = `<div>live final-turn hash: ${state.liveHashes[state.turn] || "—"}</div>`;
+}
+
+// ---- misc ---------------------------------------------------------------
+function log(msg) {
+  const el = document.getElementById("log");
+  const li = document.createElement("li");
+  li.textContent = msg;
+  el.appendChild(li);
+  el.scrollTop = el.scrollHeight;
+}
+
+window.addEventListener("DOMContentLoaded", () => {
+  initEditor();
+  initActions();
+  initReplay();
+  connect();
+});

crates/web_assets/assets/index.html

@@ -0,0 +1,82 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+  <meta charset="utf-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1" />
+  <title>Magicka VM — playable window</title>
+  <link rel="stylesheet" href="/style.css" />
+</head>
+<body>
+  <header>
+    <h1>Magicka VM</h1>
+    <div id="conn" class="badge off">disconnected</div>
+    <div id="timer" class="timer">turn —</div>
+  </header>
+
+  <main>
+    <section id="arena-panel" class="panel">
+      <h2>Arena</h2>
+      <div id="arena" class="arena" aria-label="arena grid"></div>
+      <div class="actions">
+        <div class="dpad">
+          <button data-move="0,-1">↑</button>
+          <div class="dpad-row">
+            <button data-move="-1,0">←</button>
+            <button data-move="0,1">↓</button>
+            <button data-move="1,0">→</button>
+          </div>
+        </div>
+        <div class="action-buttons">
+          <button id="btn-cast">Cast</button>
+          <button id="btn-attack">Attack</button>
+          <button id="btn-inspect">Inspect</button>
+          <button id="btn-wait">Wait</button>
+        </div>
+      </div>
+      <p class="hint">Select a target entity (click it), then Attack/Inspect. Cast uses your current program.</p>
+    </section>
+
+    <section id="editor-panel" class="panel">
+      <h2>Rune editor</h2>
+      <div id="library" class="library"></div>
+      <div id="tokens" class="tokens"></div>
+      <div class="editor-controls">
+        <select id="op-select"></select>
+        <label>a<input id="tok-a" type="number" value="0" min="0" max="255" /></label>
+        <label>b<input id="tok-b" type="number" value="0" min="0" max="255" /></label>
+        <label>c<input id="tok-c" type="number" value="0" min="0" max="255" /></label>
+        <label>imm<input id="tok-imm" type="number" value="0" /></label>
+        <button id="btn-add">Add rune</button>
+        <button id="btn-clear">Clear</button>
+        <button id="btn-save">Save program</button>
+      </div>
+      <div id="diagnostics" class="diagnostics">
+        <h3>Observed diagnostics</h3>
+        <div id="diag-body">cast or save a program to preview observed diagnostics</div>
+      </div>
+    </section>
+
+    <section id="domains-panel" class="panel">
+      <h2>Domains (observed)</h2>
+      <div id="domains" class="domains"></div>
+      <p id="redactions" class="redactions"></p>
+      <h3>Inferred</h3>
+      <ul id="inferred"></ul>
+    </section>
+
+    <section id="log-panel" class="panel">
+      <h2>Turn log</h2>
+      <ul id="log" class="log"></ul>
+      <h3>Replay</h3>
+      <div class="replay-controls">
+        <button id="btn-replay">Request replay</button>
+        <button id="btn-replay-step">Step ▶</button>
+        <span id="replay-status">no replay loaded</span>
+      </div>
+      <div id="hashes" class="hashes"></div>
+    </section>
+  </main>
+
+  <script src="/app.js"></script>
+</body>
+</html>

crates/web_assets/assets/style.css

@@ -0,0 +1,108 @@
+:root {
+  --bg: #0e1014;
+  --panel: #171a21;
+  --ink: #d7dce5;
+  --dim: #828b9c;
+  --accent: #6ad0ff;
+  --warn: #ffb454;
+  --bad: #ff6a6a;
+  --good: #7be08a;
+  --grid: #2a2f3a;
+}
+* { box-sizing: border-box; }
+body {
+  margin: 0;
+  background: var(--bg);
+  color: var(--ink);
+  font: 14px/1.4 ui-monospace, SFMono-Regular, Menlo, Consolas, monospace;
+}
+header {
+  display: flex;
+  align-items: center;
+  gap: 16px;
+  padding: 10px 16px;
+  background: #11131a;
+  border-bottom: 1px solid var(--grid);
+}
+h1 { font-size: 18px; margin: 0; color: var(--accent); }
+h2 { font-size: 14px; margin: 0 0 8px; color: var(--accent); }
+h3 { font-size: 12px; margin: 12px 0 6px; color: var(--dim); text-transform: uppercase; letter-spacing: 0.06em; }
+.badge { padding: 2px 8px; border-radius: 10px; font-size: 12px; }
+.badge.off { background: #3a1f23; color: var(--bad); }
+.badge.on { background: #1f3a26; color: var(--good); }
+.timer { margin-left: auto; color: var(--warn); }
+main {
+  display: grid;
+  grid-template-columns: 1.2fr 1.2fr 1fr;
+  grid-template-rows: auto auto;
+  gap: 12px;
+  padding: 12px;
+}
+.panel { background: var(--panel); border: 1px solid var(--grid); border-radius: 8px; padding: 12px; }
+#arena-panel { grid-row: span 2; }
+#log-panel { grid-row: span 2; }
+
+.arena {
+  display: grid;
+  gap: 2px;
+  background: var(--grid);
+  border: 1px solid var(--grid);
+  width: max-content;
+}
+.cell {
+  width: 34px; height: 34px;
+  background: #10131a;
+  display: flex; align-items: center; justify-content: center;
+  font-size: 16px; cursor: pointer; position: relative;
+}
+.cell.self { outline: 2px solid var(--accent); }
+.cell.target { outline: 2px solid var(--warn); }
+.cell.dead { opacity: 0.35; }
+.cell .hp { position: absolute; bottom: 0; right: 2px; font-size: 9px; color: var(--dim); }
+
+.actions { display: flex; gap: 24px; margin-top: 12px; align-items: center; }
+.dpad { display: flex; flex-direction: column; align-items: center; gap: 2px; }
+.dpad-row { display: flex; gap: 2px; }
+button {
+  background: #222733; color: var(--ink); border: 1px solid var(--grid);
+  border-radius: 5px; padding: 6px 10px; cursor: pointer; font: inherit;
+}
+button:hover { border-color: var(--accent); }
+button:disabled { opacity: 0.4; cursor: not-allowed; }
+.action-buttons { display: flex; flex-wrap: wrap; gap: 6px; }
+.hint { color: var(--dim); font-size: 12px; }
+
+.library { display: flex; gap: 6px; margin-bottom: 8px; }
+.slot { border: 1px dashed var(--grid); border-radius: 5px; padding: 4px 8px; cursor: pointer; color: var(--dim); }
+.slot.active { border-color: var(--accent); color: var(--accent); }
+.tokens { display: flex; flex-wrap: wrap; gap: 4px; min-height: 30px; padding: 6px; background: #10131a; border-radius: 5px; }
+.token { background: #232a36; border: 1px solid var(--grid); border-radius: 4px; padding: 2px 6px; font-size: 12px; cursor: pointer; }
+.token:hover { border-color: var(--bad); }
+.editor-controls { display: flex; flex-wrap: wrap; gap: 6px; align-items: center; margin-top: 8px; }
+.editor-controls label { display: flex; flex-direction: column; font-size: 10px; color: var(--dim); }
+.editor-controls input { width: 64px; background: #10131a; border: 1px solid var(--grid); color: var(--ink); border-radius: 4px; padding: 3px; }
+.editor-controls select { background: #10131a; border: 1px solid var(--grid); color: var(--ink); border-radius: 4px; padding: 4px; }
+
+.diagnostics { margin-top: 12px; background: #10131a; border-radius: 5px; padding: 8px; }
+.diag-row { margin: 2px 0; }
+.diag-key { color: var(--dim); }
+.warn { color: var(--warn); }
+
+.domains { display: grid; grid-template-columns: 1fr 1fr; gap: 6px; }
+.domain { background: #10131a; border: 1px solid var(--grid); border-radius: 5px; padding: 6px; }
+.domain .name { color: var(--accent); font-size: 12px; }
+.lane { display: inline-block; min-width: 40px; text-align: right; padding: 1px 4px; margin: 1px; border-radius: 3px; font-size: 11px; }
+.lane.known { background: #15251a; color: var(--good); }
+.lane.newly_observed { background: #2a2410; color: var(--warn); }
+.lane.unknown { background: #25151a; color: var(--dim); }
+.lane.suspected { background: #1a1a2a; color: #9aa0ff; }
+.lane.contradicted { background: #2a1525; color: #ff9ae0; }
+.redactions { color: var(--bad); font-size: 12px; }
+
+.log { list-style: none; margin: 0; padding: 0; max-height: 320px; overflow-y: auto; }
+.log li { padding: 2px 0; border-bottom: 1px solid #1c2029; font-size: 12px; }
+.replay-controls { display: flex; gap: 6px; align-items: center; flex-wrap: wrap; }
+.hashes { margin-top: 8px; font-size: 11px; color: var(--dim); word-break: break-all; }
+.hash-ok { color: var(--good); }
+.hash-bad { color: var(--bad); }
+ul#inferred { margin: 0; padding-left: 16px; color: var(--dim); font-size: 12px; }

crates/web_assets/src/lib.rs

@@ -0,0 +1,36 @@
+//! `web_assets` — the browser client's static files, embedded at compile time
+//! so the server ships as a single binary with no runtime filesystem
+//! dependency. The actual HTML/CSS/JS live under `assets/`.
+
+pub const INDEX_HTML: &str = include_str!("../assets/index.html");
+pub const STYLE_CSS: &str = include_str!("../assets/style.css");
+pub const APP_JS: &str = include_str!("../assets/app.js");
+
+/// A served asset: its bytes and MIME type.
+pub struct Asset {
+    pub body: &'static str,
+    pub content_type: &'static str,
+}
+
+/// Resolve a request path to a static asset. `/` maps to the client shell.
+pub fn resolve(path: &str) -> Option<Asset> {
+    match path {
+        "/" | "/index.html" => Some(Asset { body: INDEX_HTML, content_type: "text/html; charset=utf-8" }),
+        "/style.css" => Some(Asset { body: STYLE_CSS, content_type: "text/css; charset=utf-8" }),
+        "/app.js" => Some(Asset { body: APP_JS, content_type: "application/javascript; charset=utf-8" }),
+        _ => None,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn shell_and_assets_resolve() {
+        assert!(resolve("/").is_some());
+        assert!(resolve("/app.js").is_some());
+        assert!(resolve("/style.css").is_some());
+        assert!(resolve("/nope").is_none());
+    }
+}

crates/web_client/Cargo.toml

@@ -0,0 +1,8 @@
+[package]
+name = "web_client"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+web_assets = { path = "../web_assets" }

crates/web_client/src/lib.rs

@@ -0,0 +1,54 @@
+//! `web_client` — the client delivery layer. It owns *how* the embedded
+//! [`web_assets`] reach the browser (the HTTP response framing), keeping the
+//! raw asset bytes (`web_assets`) separate from delivery concerns. The server
+//! depends on this crate, not on `web_assets` directly.
+
+pub use web_assets::{resolve, Asset};
+
+/// Build a complete HTTP/1.1 response for a static GET path. Returns `None`
+/// for unknown paths so the caller can emit a 404.
+pub fn http_response(path: &str) -> Option<Vec<u8>> {
+    let asset = resolve(path)?;
+    let body = asset.body.as_bytes();
+    let mut out = Vec::with_capacity(body.len() + 128);
+    let header = format!(
+        "HTTP/1.1 200 OK\r\nContent-Type: {}\r\nContent-Length: {}\r\nCache-Control: no-cache\r\nConnection: close\r\n\r\n",
+        asset.content_type,
+        body.len()
+    );
+    out.extend_from_slice(header.as_bytes());
+    out.extend_from_slice(body);
+    Some(out)
+}
+
+/// The canonical 404 response.
+pub fn not_found() -> Vec<u8> {
+    let body = b"404 not found";
+    let mut out = Vec::new();
+    let header = format!(
+        "HTTP/1.1 404 Not Found\r\nContent-Type: text/plain\r\nContent-Length: {}\r\nConnection: close\r\n\r\n",
+        body.len()
+    );
+    out.extend_from_slice(header.as_bytes());
+    out.extend_from_slice(body);
+    out
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn serves_shell() {
+        let r = http_response("/").unwrap();
+        let s = String::from_utf8_lossy(&r);
+        assert!(s.starts_with("HTTP/1.1 200 OK"));
+        assert!(s.contains("text/html"));
+        assert!(s.contains("Magicka VM"));
+    }
+
+    #[test]
+    fn unknown_path_is_none() {
+        assert!(http_response("/secret").is_none());
+    }
+}

crates/web_tests/Cargo.toml

@@ -0,0 +1,12 @@
+[package]
+name = "web_tests"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+protocol = { path = "../protocol" }
+game_runtime = { path = "../game_runtime" }
+world_model = { path = "../world_model" }
+rune_ir = { path = "../rune_ir" }
+server = { path = "../server" }

crates/web_tests/e2e/package.json

@@ -0,0 +1,12 @@
+{
+  "name": "magicka-web-e2e",
+  "version": "0.1.0",
+  "private": true,
+  "description": "Playwright rendered-browser E2E for the Magicka VM web game.",
+  "scripts": {
+    "test": "playwright test"
+  },
+  "devDependencies": {
+    "@playwright/test": "^1.40.0"
+  }
+}

crates/web_tests/e2e/playwright.config.js

@@ -0,0 +1,28 @@
+// Playwright config for the rendered-browser E2E layer. This complements the
+// headless Rust protocol E2E in `crates/web_tests/tests/e2e.rs`: here a real
+// Chromium drives the actual DOM client. It boots the real server (short turns)
+// via `webServer` so `npm test` is self-contained.
+//
+// Requires Node + `npx playwright install chromium`. The Rust CI gates do not
+// depend on this; it is the optional rendered-browser proof.
+
+const { defineConfig } = require("@playwright/test");
+
+module.exports = defineConfig({
+  testDir: "./specs",
+  timeout: 30000,
+  expect: { timeout: 10000 },
+  use: {
+    baseURL: "http://127.0.0.1:8099",
+    headless: true,
+  },
+  webServer: {
+    // Build once, then run the server with fast turns for E2E.
+    command: "cargo run --release -p server --bin magicka-server",
+    cwd: "../../..",
+    env: { MAGICKA_ADDR: "127.0.0.1:8099", MAGICKA_TURN_MS: "1500" },
+    url: "http://127.0.0.1:8099/",
+    reuseExistingServer: true,
+    timeout: 120000,
+  },
+});

crates/web_tests/e2e/specs/play.spec.js

@@ -0,0 +1,43 @@
+// Rendered-browser E2E: a real Chromium joins a match, the turn timer runs, the
+// player edits + saves a rune program, casts, sees filtered results, and replays
+// the match — verifying the recorded replay hashes match what was seen live.
+//
+// This is the browser layer of plan2.md Phase H ("Playwright end-to-end tests").
+
+const { test, expect } = require("@playwright/test");
+
+test("a player can join, cast, and replay a browser match", async ({ page }) => {
+  await page.goto("/");
+
+  // Connects to the authoritative server.
+  await expect(page.locator("#conn")).toHaveText("connected", { timeout: 10000 });
+
+  // The turn timer is running (header shows a turn + countdown).
+  await expect(page.locator("#timer")).toContainText("turn", { timeout: 10000 });
+
+  // The arena rendered with the player's marker.
+  await expect(page.locator(".cell.self")).toHaveCount(1, { timeout: 10000 });
+
+  // Domains panel shows the hidden-state redaction notice (visibility layer).
+  await expect(page.locator("#redactions")).toContainText("withheld", { timeout: 10000 });
+
+  // Edit a rune program: add a couple of runes and save.
+  await page.fill("#tok-a", "1");
+  await page.fill("#tok-b", "2");
+  await page.click("#btn-add");
+  await page.click("#btn-add");
+  await page.click("#btn-save");
+  // Observed diagnostics appear (names/counts only).
+  await expect(page.locator("#diag-body")).toContainText("known reads", { timeout: 10000 });
+
+  // Cast and wait for a resolved-turn log line.
+  await page.click("#btn-cast");
+  await expect(page.locator("#log")).toContainText("cast a rune program", { timeout: 15000 });
+
+  // Request and step the replay; the client verifies hashes vs. what it saw live.
+  await page.click("#btn-replay");
+  await expect(page.locator("#replay-status")).toContainText("replay loaded", { timeout: 10000 });
+  await page.click("#btn-replay-step");
+  // A matching hash line is shown (no MISMATCH).
+  await expect(page.locator(".hash-bad")).toHaveCount(0);
+});

crates/web_tests/src/lib.rs

@@ -0,0 +1,170 @@
+//! `web_tests` — a tiny, dependency-free WebSocket *client* used to drive the
+//! real server over a real socket in integration tests. It performs the HTTP
+//! upgrade, masks client frames (as RFC 6455 requires), and reads server
+//! frames. This is the harness behind the Phase H web CI gates.
+
+use std::io::{self, BufRead, BufReader, Read, Write};
+use std::net::TcpStream;
+use std::time::Duration;
+
+use protocol::{ClientMessage, ServerMessage};
+
+/// A blocking WebSocket client connection to the test server.
+pub struct WsClient {
+    stream: TcpStream,
+    reader: BufReader<TcpStream>,
+}
+
+impl WsClient {
+    /// Connect, upgrade to WebSocket, and verify the handshake.
+    pub fn connect(addr: &str) -> io::Result<WsClient> {
+        let stream = TcpStream::connect(addr)?;
+        stream.set_read_timeout(Some(Duration::from_secs(5)))?;
+        stream.set_nodelay(true).ok();
+        let mut reader = BufReader::new(stream.try_clone()?);
+
+        // A fixed client key keeps the handshake assertion deterministic.
+        let key = "dGhlIHNhbXBsZSBub25jZQ==";
+        let mut s = stream.try_clone()?;
+        let req = format!(
+            "GET /ws HTTP/1.1\r\nHost: localhost\r\nUpgrade: websocket\r\n\
+             Connection: Upgrade\r\nSec-WebSocket-Key: {key}\r\n\
+             Sec-WebSocket-Version: 13\r\n\r\n"
+        );
+        s.write_all(req.as_bytes())?;
+        s.flush()?;
+
+        // Read the response head.
+        let mut status = String::new();
+        reader.read_line(&mut status)?;
+        if !status.contains("101") {
+            return Err(io::Error::new(io::ErrorKind::Other, format!("no upgrade: {status:?}")));
+        }
+        let mut saw_accept = false;
+        loop {
+            let mut line = String::new();
+            let n = reader.read_line(&mut line)?;
+            if n == 0 || line.trim_end().is_empty() {
+                break;
+            }
+            if line.to_ascii_lowercase().starts_with("sec-websocket-accept:") {
+                let got = line.split(':').nth(1).unwrap_or("").trim();
+                // Expected accept for the canonical key above.
+                if got == "s3pPLMBiTxaQ9kYGzzhZRbK+xOo=" {
+                    saw_accept = true;
+                }
+            }
+        }
+        if !saw_accept {
+            return Err(io::Error::new(io::ErrorKind::Other, "bad Sec-WebSocket-Accept"));
+        }
+        Ok(WsClient { stream, reader })
+    }
+
+    /// Send a typed client message.
+    pub fn send(&mut self, msg: &ClientMessage) -> io::Result<()> {
+        self.send_raw_text(&msg.encode())
+    }
+
+    /// Send arbitrary text as a masked frame (used by fuzz tests).
+    pub fn send_raw_text(&mut self, text: &str) -> io::Result<()> {
+        self.write_masked(0x1, text.as_bytes())
+    }
+
+    /// Send arbitrary bytes as a masked binary frame (fuzz transport).
+    pub fn send_raw_bytes(&mut self, bytes: &[u8]) -> io::Result<()> {
+        self.write_masked(0x2, bytes)
+    }
+
+    fn write_masked(&mut self, opcode: u8, payload: &[u8]) -> io::Result<()> {
+        let mask = [0x12u8, 0x34, 0x56, 0x78];
+        let mut frame = vec![0x80 | opcode];
+        let len = payload.len();
+        if len < 126 {
+            frame.push(0x80 | len as u8);
+        } else if len < 65536 {
+            frame.push(0x80 | 126);
+            frame.extend_from_slice(&(len as u16).to_be_bytes());
+        } else {
+            frame.push(0x80 | 127);
+            frame.extend_from_slice(&(len as u64).to_be_bytes());
+        }
+        frame.extend_from_slice(&mask);
+        for (i, &b) in payload.iter().enumerate() {
+            frame.push(b ^ mask[i % 4]);
+        }
+        self.stream.write_all(&frame)?;
+        self.stream.flush()
+    }
+
+    /// Read one server text frame and decode it. Skips control frames.
+    pub fn recv(&mut self) -> io::Result<ServerMessage> {
+        let text = self.recv_text()?;
+        ServerMessage::decode(&text)
+            .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, format!("decode: {e}")))
+    }
+
+    /// Read one server text frame (raw).
+    pub fn recv_text(&mut self) -> io::Result<String> {
+        loop {
+            let mut hdr = [0u8; 2];
+            self.reader.read_exact(&mut hdr)?;
+            let opcode = hdr[0] & 0x0f;
+            let masked = hdr[1] & 0x80 != 0;
+            let len7 = (hdr[1] & 0x7f) as usize;
+            let len = match len7 {
+                126 => {
+                    let mut b = [0u8; 2];
+                    self.reader.read_exact(&mut b)?;
+                    u16::from_be_bytes(b) as usize
+                }
+                127 => {
+                    let mut b = [0u8; 8];
+                    self.reader.read_exact(&mut b)?;
+                    u64::from_be_bytes(b) as usize
+                }
+                n => n,
+            };
+            // Server frames are not masked, but tolerate it.
+            let mask = if masked {
+                let mut m = [0u8; 4];
+                self.reader.read_exact(&mut m)?;
+                Some(m)
+            } else {
+                None
+            };
+            let mut payload = vec![0u8; len];
+            self.reader.read_exact(&mut payload)?;
+            if let Some(m) = mask {
+                for (i, b) in payload.iter_mut().enumerate() {
+                    *b ^= m[i % 4];
+                }
+            }
+            match opcode {
+                0x1 | 0x2 => return Ok(String::from_utf8_lossy(&payload).into_owned()),
+                0x8 => return Err(io::Error::new(io::ErrorKind::ConnectionAborted, "closed")),
+                _ => continue, // ping/pong/continuation
+            }
+        }
+    }
+
+    /// Receive until a predicate matches, returning that message. Bounded so a
+    /// test never hangs.
+    pub fn recv_until<F: Fn(&ServerMessage) -> bool>(&mut self, pred: F) -> io::Result<ServerMessage> {
+        for _ in 0..256 {
+            let m = self.recv()?;
+            if pred(&m) {
+                return Ok(m);
+            }
+        }
+        Err(io::Error::new(io::ErrorKind::Other, "predicate never matched"))
+    }
+}
+
+/// Start a fresh server instance on an ephemeral port for a test, returning the
+/// address string. Each call binds a new port so tests are isolated.
+pub fn spawn_test_server(turn_ms: u64) -> String {
+    let addr = server::serve("127.0.0.1:0", server::Config { turn_ms })
+        .expect("bind test server");
+    format!("127.0.0.1:{}", addr.port())
+}

crates/web_tests/tests/determinism.rs

@@ -0,0 +1,64 @@
+//! Phase H gate: replay determinism. The web CI minimum is "1,000 simulated
+//! matches" with "0 replay hash mismatches". A match is a pure function of its
+//! seed, roster, and ordered inputs, so re-running the recorded inputs must
+//! reproduce the final hash exactly.
+
+use game_runtime::{replay, run_scripted, solo_roster, duel_roster};
+use protocol::Action;
+use world_model::Rng;
+
+/// Build a varied but deterministic script for a given seed.
+fn script(seed: u64) -> Vec<Vec<(u32, Action)>> {
+    let mut rng = Rng::derive(seed, "script");
+    let mut turns = Vec::new();
+    for _ in 0..6 {
+        let mut subs = Vec::new();
+        for pid in 1..=2u32 {
+            let a = match rng.below(5) {
+                0 => Action::Move { dx: rng.range_i64(-1, 1) as i32, dy: rng.range_i64(-1, 1) as i32 },
+                1 => Action::Cast,
+                2 => Action::Attack { target: if pid == 1 { 2 } else { 1 } },
+                3 => Action::Inspect { target: if pid == 1 { 2 } else { 1 } },
+                _ => Action::Wait,
+            };
+            subs.push((pid, a));
+        }
+        turns.push(subs);
+    }
+    turns
+}
+
+#[test]
+fn one_thousand_matches_replay_with_zero_drift() {
+    let mut mismatches = 0u32;
+    for seed in 0..1000u64 {
+        let roster = if seed % 2 == 0 {
+            solo_roster("p")
+        } else {
+            duel_roster("a", "b")
+        };
+        let scripts = script(seed);
+        let (m, log) = run_scripted(seed, &roster, &scripts);
+        let again = replay(seed, &roster, &log.turns);
+        if m.replay.final_hash != again.final_hash {
+            mismatches += 1;
+        }
+        // Per-turn hashes must also agree.
+        for (a, b) in m.replay.turns.iter().zip(again.turns.iter()) {
+            if a.turn_hash != b.turn_hash {
+                mismatches += 1;
+            }
+        }
+    }
+    assert_eq!(mismatches, 0, "replay hash mismatches across 1000 matches");
+}
+
+#[test]
+fn re_executing_same_seed_twice_is_identical() {
+    for seed in [1u64, 7, 99, 12345, 0xdead_beef] {
+        let roster = solo_roster("p");
+        let (a, _) = run_scripted(seed, &roster, &script(seed));
+        let (b, _) = run_scripted(seed, &roster, &script(seed));
+        assert_eq!(a.replay.final_hash, b.replay.final_hash, "seed {seed}");
+    }
+}

crates/web_tests/tests/e2e.rs

@@ -0,0 +1,145 @@
+//! Phase H gate: end-to-end matches over the real server, real sockets, real
+//! protocol. The web CI minimum names "100 browser E2E matches"; a headless
+//! browser is not available in this CI, so this drives the full
+//! HTTP+WebSocket+protocol+runtime path *headlessly* (it is protocol-level
+//! E2E, not a rendered browser — the Playwright harness under `e2e/` covers the
+//! rendered browser when Node is present). It asserts: a player can join, the
+//! turn timer drives resolution, casts resolve through the runtime, results
+//! return as filtered observations, and the recorded replay reproduces the live
+//! per-turn hashes (0 replay hash mismatches).
+
+use protocol::{Action, ClientMessage, MatchId, RuneTokenWire, ServerMessage};
+use web_tests::{spawn_test_server, WsClient};
+
+struct Played {
+    match_id: MatchId,
+    live_hashes: Vec<(u64, String)>,
+    replay_hashes: Vec<(u64, String)>,
+    final_hash: String,
+    saw_filtered_snapshot: bool,
+}
+
+/// Play a solo match for `n_turns`, then pull the recorded replay back.
+fn play_solo(addr: &str, n_turns: u64, program: &[RuneTokenWire]) -> std::io::Result<Played> {
+    let mut c = WsClient::connect(addr)?;
+    c.send(&ClientMessage::JoinMatch { name: "e2e".into(), match_id: None })?;
+    let (match_id, _player) = match c.recv_until(|m| matches!(m, ServerMessage::MatchState { .. }))? {
+        ServerMessage::MatchState { match_id, player_id, snapshot, .. } => {
+            // Filtered snapshot sanity: hidden lanes are redacted.
+            assert!(snapshot.hidden_state_redactions > 0, "no redactions in snapshot");
+            (match_id, player_id)
+        }
+        _ => unreachable!(),
+    };
+
+    if !program.is_empty() {
+        c.send(&ClientMessage::EditRuneProgram { tokens: program.to_vec() })?;
+    }
+
+    let mut live_hashes = Vec::new();
+    let mut saw_filtered = false;
+    let mut live_turn = 0u64;
+    for _ in 0..n_turns {
+        c.send(&ClientMessage::SubmitTurn { turn: live_turn, action: Action::Cast })?;
+        let expected = live_turn + 1;
+        let resolved = c.recv_until(|m| {
+            matches!(m, ServerMessage::TurnResolved { turn, .. } if *turn == expected)
+        })?;
+        if let ServerMessage::TurnResolved { turn, runtime_hash, snapshot, .. } = resolved {
+            // Results come back as filtered observations.
+            if snapshot.hidden_state_redactions > 0 {
+                saw_filtered = true;
+            }
+            live_hashes.push((turn, runtime_hash));
+            live_turn = turn;
+        }
+    }
+
+    // Pull the replay back and collect its per-turn hashes.
+    c.send(&ClientMessage::RequestReplay { match_id })?;
+    let mut replay_hashes = Vec::new();
+    let mut final_hash = String::new();
+    let mut got = 0u32;
+    let mut total = 1u32;
+    while got < total {
+        match c.recv_until(|m| matches!(m, ServerMessage::ReplayChunk { .. }))? {
+            ServerMessage::ReplayChunk { total: t, turns, final_hash: fh, .. } => {
+                total = t.max(1);
+                got += 1;
+                final_hash = fh;
+                for rt in turns {
+                    replay_hashes.push((rt.turn, rt.runtime_hash));
+                }
+            }
+            _ => break,
+        }
+    }
+
+    Ok(Played { match_id, live_hashes, replay_hashes, final_hash, saw_filtered_snapshot: saw_filtered })
+}
+
+#[test]
+fn single_match_full_playthrough() {
+    let addr = spawn_test_server(15);
+    let program = vec![
+        RuneTokenWire { op: 0, a: 1, b: 2, c: 3, imm: 5 },
+        RuneTokenWire { op: 5, a: 2, b: 4, c: 1, imm: -3 },
+        RuneTokenWire { op: 8, a: 0, b: 6, c: 2, imm: 11 },
+    ];
+    let played = play_solo(&addr, 4, &program).expect("play");
+    assert_ne!(played.match_id.0, 0, "a match id was assigned");
+    assert_eq!(played.live_hashes.len(), 4, "all turns resolved");
+    assert!(played.saw_filtered_snapshot, "results returned as filtered observations");
+    assert!(!played.final_hash.is_empty());
+
+    // Replay must reproduce every live per-turn hash (0 mismatches).
+    for (turn, live) in &played.live_hashes {
+        let found = played.replay_hashes.iter().find(|(t, _)| t == turn);
+        assert!(found.is_some(), "replay missing turn {turn}");
+        assert_eq!(&found.unwrap().1, live, "replay hash mismatch at turn {turn}");
+    }
+}
+
+#[test]
+fn inspection_returns_observed_diagnostics_only() {
+    let addr = spawn_test_server(50);
+    let mut c = WsClient::connect(&addr).unwrap();
+    c.send(&ClientMessage::JoinMatch { name: "inspector".into(), match_id: None }).unwrap();
+    c.recv_until(|m| matches!(m, ServerMessage::MatchState { .. })).unwrap();
+    // Inspect the dummy (entity id 2).
+    c.send(&ClientMessage::InspectTarget { target: 2 }).unwrap();
+    let obs = c.recv_until(|m| matches!(m, ServerMessage::ObservationResult { .. })).unwrap();
+    if let ServerMessage::ObservationResult { diagnostics, .. } = obs {
+        // Diagnostics are observed names/counts — never guaranteed truth.
+        // unknown_listeners is a count; reads/writes are domain names.
+        for s in diagnostics.known_reads.iter().chain(diagnostics.known_writes.iter()) {
+            assert!(s.chars().any(|ch| ch.is_alphabetic()));
+        }
+    } else {
+        panic!("expected ObservationResult");
+    }
+}
+
+#[test]
+fn one_hundred_e2e_matches_zero_hash_mismatches() {
+    // Single short-turn server; 100 independent matches over real sockets.
+    let addr = spawn_test_server(8);
+    let program = vec![
+        RuneTokenWire { op: 1, a: 3, b: 1, c: 2, imm: 7 },
+        RuneTokenWire { op: 6, a: 0, b: 5, c: 3, imm: -9 },
+    ];
+    let mut mismatches = 0u32;
+    let mut completed = 0u32;
+    for _ in 0..100 {
+        let played = play_solo(&addr, 3, &program).expect("e2e match");
+        completed += 1;
+        for (turn, live) in &played.live_hashes {
+            match played.replay_hashes.iter().find(|(t, _)| t == turn) {
+                Some((_, rh)) if rh == live => {}
+                _ => mismatches += 1,
+            }
+        }
+    }
+    assert_eq!(completed, 100, "all 100 matches completed");
+    assert_eq!(mismatches, 0, "replay hash mismatches across 100 e2e matches");
+}

crates/web_tests/tests/fuzz.rs

@@ -0,0 +1,84 @@
+//! Phase H gate: protocol fuzzing. The web CI minimum is "10,000 protocol fuzz
+//! cases" with "0 server panics". Decoding is total, so every byte string must
+//! yield `Ok` or `Err` — never an unwind. A sample is also fired at a live
+//! server to prove a malformed packet cannot bring it down.
+
+use protocol::{json, ClientMessage};
+use web_tests::{spawn_test_server, WsClient};
+use world_model::Rng;
+
+/// Generate a pseudo-random byte string from a seed, with a bias toward
+/// JSON-ish characters so the parser's deeper paths are exercised.
+fn fuzz_bytes(seed: u64) -> Vec<u8> {
+    let mut rng = Rng::derive(seed, "fuzz");
+    let alphabet = b"{}[]\":,0123456789tfnuelavabcdef.- \\/\n\t";
+    let len = rng.below(80);
+    (0..len)
+        .map(|_| {
+            if rng.chance(0.85) {
+                alphabet[rng.below(alphabet.len())]
+            } else {
+                rng.next_u64() as u8
+            }
+        })
+        .collect()
+}
+
+#[test]
+fn ten_thousand_fuzz_cases_never_panic() {
+    let mut ok = 0u64;
+    let mut err = 0u64;
+    for seed in 0..10_000u64 {
+        let bytes = fuzz_bytes(seed);
+        let text = String::from_utf8_lossy(&bytes);
+        // Raw JSON parse must be total.
+        let _ = json::parse(&text);
+        // Full client-message decode must be total.
+        match ClientMessage::decode(&text) {
+            Ok(_) => ok += 1,
+            Err(_) => err += 1,
+        }
+    }
+    // The point is the absence of a panic; both counters are just evidence the
+    // loop ran to completion.
+    assert_eq!(ok + err, 10_000);
+}
+
+#[test]
+fn structured_but_invalid_messages_are_rejected_not_panicked() {
+    let cases = [
+        "{}",
+        "{\"v\":1}",
+        "{\"v\":2,\"type\":\"Ping\",\"body\":{}}",         // wrong version
+        "{\"v\":1,\"type\":\"Nope\",\"body\":{}}",          // unknown type
+        "{\"v\":1,\"type\":\"SubmitTurn\",\"body\":{}}",    // missing fields
+        "{\"v\":1,\"type\":\"JoinMatch\",\"body\":{\"name\":5}}", // wrong type
+    ];
+    for c in cases {
+        assert!(ClientMessage::decode(c).is_err(), "should reject: {c}");
+    }
+}
+
+#[test]
+fn live_server_survives_malformed_packets() {
+    let addr = spawn_test_server(30);
+    let mut c = WsClient::connect(&addr).expect("connect");
+    // Fire a burst of garbage frames (kept modest so the join reply is not
+    // starved behind a flood of rejection reports; decode breadth is covered by
+    // the 10k case test above).
+    for seed in 0..40u64 {
+        let bytes = fuzz_bytes(seed);
+        let _ = c.send_raw_bytes(&bytes);
+    }
+    // Also send raw garbage text.
+    for s in ["", "{", "garbage", "{\"v\":1,\"type\":\"X\",\"body\":1}"] {
+        let _ = c.send_raw_text(s);
+    }
+    // The server must still be alive and respond to a valid join.
+    c.send(&ClientMessage::JoinMatch { name: "after-fuzz".into(), match_id: None })
+        .expect("send join");
+    let m = c
+        .recv_until(|m| matches!(m, protocol::ServerMessage::MatchState { .. }))
+        .expect("server still serving after fuzz");
+    assert!(matches!(m, protocol::ServerMessage::MatchState { .. }));
+}

crates/web_tests/tests/resilience.rs

@@ -0,0 +1,94 @@
+//! Phase H gates: disconnect/reconnect and timer edges. A dropped connection
+//! must not corrupt a match, and a late or wrong-turn submission must be
+//! rejected deterministically.
+
+use protocol::{Action, ClientMessage, ServerMessage};
+use web_tests::{spawn_test_server, WsClient};
+
+#[test]
+fn wrong_turn_submission_is_rejected_deterministically() {
+    let addr = spawn_test_server(2000); // long turn so we control timing
+    let mut c = WsClient::connect(&addr).unwrap();
+    c.send(&ClientMessage::JoinMatch { name: "timer".into(), match_id: None }).unwrap();
+    c.recv_until(|m| matches!(m, ServerMessage::MatchState { .. })).unwrap();
+
+    // Submit for a turn that is not live.
+    c.send(&ClientMessage::SubmitTurn { turn: 999, action: Action::Wait }).unwrap();
+    let report = c.recv_until(|m| matches!(m, ServerMessage::ValidationReport { .. })).unwrap();
+    match report {
+        ServerMessage::ValidationReport { accepted, detail, .. } => {
+            assert!(!accepted, "wrong-turn submission should be rejected");
+            assert!(detail.contains("wrong turn"), "reason: {detail}");
+        }
+        _ => unreachable!(),
+    }
+
+    // A correct, in-time submission is accepted.
+    c.send(&ClientMessage::SubmitTurn { turn: 0, action: Action::Wait }).unwrap();
+    let ok = c.recv_until(|m| matches!(m, ServerMessage::ValidationReport { .. })).unwrap();
+    match ok {
+        ServerMessage::ValidationReport { accepted, .. } => assert!(accepted),
+        _ => unreachable!(),
+    }
+}
+
+#[test]
+fn disconnect_does_not_corrupt_an_ongoing_duel() {
+    let addr = spawn_test_server(10);
+    // Two players share a duel match by id.
+    let mut a = WsClient::connect(&addr).unwrap();
+    a.send(&ClientMessage::JoinMatch { name: "a".into(), match_id: Some(protocol::MatchId(7)) })
+        .unwrap();
+    let _ = a.recv_until(|m| matches!(m, ServerMessage::MatchState { .. })).unwrap();
+
+    let mut b = WsClient::connect(&addr).unwrap();
+    b.send(&ClientMessage::JoinMatch { name: "b".into(), match_id: Some(protocol::MatchId(7)) })
+        .unwrap();
+    let _ = b.recv_until(|m| matches!(m, ServerMessage::MatchState { .. })).unwrap();
+
+    // Player A plays a couple of turns.
+    let mut live = 0u64;
+    for _ in 0..2 {
+        a.send(&ClientMessage::SubmitTurn { turn: live, action: Action::Cast }).unwrap();
+        let r = a
+            .recv_until(|m| matches!(m, ServerMessage::TurnResolved { turn, .. } if *turn == live + 1))
+            .unwrap();
+        if let ServerMessage::TurnResolved { turn, .. } = r {
+            live = turn;
+        }
+    }
+
+    // Player B disconnects abruptly.
+    drop(b);
+
+    // The match continues for A without corruption: more turns still resolve.
+    for _ in 0..2 {
+        a.send(&ClientMessage::SubmitTurn { turn: live, action: Action::Cast }).unwrap();
+        let r = a
+            .recv_until(|m| matches!(m, ServerMessage::TurnResolved { turn, .. } if *turn == live + 1))
+            .unwrap();
+        if let ServerMessage::TurnResolved { turn, runtime_hash, .. } = r {
+            assert!(!runtime_hash.is_empty());
+            live = turn;
+        }
+    }
+    assert!(live >= 4, "match advanced past a mid-match disconnect");
+}
+
+#[test]
+fn reconnect_into_open_duel_slot_succeeds() {
+    let addr = spawn_test_server(50);
+    let mid = protocol::MatchId(21);
+
+    let mut a = WsClient::connect(&addr).unwrap();
+    a.send(&ClientMessage::JoinMatch { name: "host".into(), match_id: Some(mid) }).unwrap();
+    a.recv_until(|m| matches!(m, ServerMessage::MatchState { .. })).unwrap();
+
+    // A second player takes the open slot.
+    let mut b = WsClient::connect(&addr).unwrap();
+    b.send(&ClientMessage::JoinMatch { name: "guest".into(), match_id: Some(mid) }).unwrap();
+    let ms = b.recv_until(|m| matches!(m, ServerMessage::MatchState { .. })).unwrap();
+    if let ServerMessage::MatchState { player_id, .. } = ms {
+        assert_eq!(player_id.0, 2, "second human should take slot 2");
+    }
+}

crates/web_tests/tests/visibility.rs

@@ -0,0 +1,74 @@
+//! Phase H gate: "0 hidden-state leaks". The hidden ground truth must never
+//! cross the wire. These tests check the invariant both in-process (the
+//! authoritative match) and over the live socket (the serialized bytes).
+
+use game_runtime::{solo_roster, Match};
+use protocol::{Action, ClientMessage, MatchId, ServerMessage};
+use web_tests::{spawn_test_server, WsClient};
+use world_model::{HIDDEN_LANES, LANES, NUM_DOMAINS};
+
+#[test]
+fn visible_snapshot_redacts_all_hidden_lanes() {
+    let mut m = Match::new(MatchId(1), 4242, solo_roster("dev"));
+    // Mask half of the observed lanes so redaction is non-trivial too.
+    for d in 0..NUM_DOMAINS {
+        for l in 0..LANES {
+            if (d + l) % 2 == 0 {
+                m.world.observation_state.visible[d][l] = false;
+            }
+        }
+    }
+    // The generator may already mask lanes; count actual non-visible lanes.
+    let masked: u32 = (0..NUM_DOMAINS)
+        .flat_map(|d| (0..LANES).map(move |l| (d, l)))
+        .filter(|&(d, l)| !m.world.observation_state.visible[d][l])
+        .count() as u32;
+    m.set_program(1, vec![protocol::RuneTokenWire { op: 0, a: 1, b: 2, c: 3, imm: 1 }]);
+    m.resolve_turn(&[(1, Action::Cast)]);
+
+    let snap = m.visible_for(1);
+    // Every masked observed lane is None.
+    for vd in &snap.observed_domains {
+        assert_eq!(vd.observed.len(), LANES);
+        for (l, o) in vd.observed.iter().enumerate() {
+            let visible = m.world.observation_state.visible[vd.index as usize][l];
+            assert_eq!(o.is_some(), visible, "lane visibility/value mismatch");
+        }
+    }
+    // Redaction count accounts for every masked observed lane and every hidden
+    // lane in the world.
+    let expected = masked + (NUM_DOMAINS * HIDDEN_LANES) as u32;
+    assert_eq!(snap.hidden_state_redactions, expected);
+
+    // The serialized snapshot must not carry a "hidden" key at all.
+    let wire = snap.to_json().to_compact();
+    assert!(!wire.contains("\"hidden\""), "serialized snapshot mentions hidden state");
+}
+
+#[test]
+fn no_server_message_over_the_wire_carries_hidden_keys() {
+    let addr = spawn_test_server(12);
+    let mut c = WsClient::connect(&addr).unwrap();
+    c.send(&ClientMessage::JoinMatch { name: "leak-check".into(), match_id: None }).unwrap();
+
+    let mut checked = 0;
+    // Drive a few turns and scan every raw frame the server emits.
+    let mut live_turn = 0u64;
+    for _ in 0..6 {
+        c.send(&ClientMessage::SubmitTurn { turn: live_turn, action: Action::Cast }).unwrap();
+        // Read several frames; scan each.
+        for _ in 0..4 {
+            let raw = match c.recv_text() {
+                Ok(r) => r,
+                Err(_) => break,
+            };
+            checked += 1;
+            assert!(!raw.contains("\"hidden\""), "raw frame leaked hidden key: {raw}");
+            // Track turn progression from resolved frames.
+            if let Ok(ServerMessage::TurnResolved { turn, .. }) = ServerMessage::decode(&raw) {
+                live_turn = turn;
+            }
+        }
+    }
+    assert!(checked > 0, "no frames scanned");
+}

crates/world_model/src/world.rs

@@ -262,4 +262,45 @@ impl WorldDelta {
         }
         h.finish()
     }
+
+    /// Serialize the full delta to one line of whitespace-separated integers.
+    /// Round-trips [`hash`] exactly (used as attestation evidence).
+    pub fn serialize(&self) -> String {
+        let mut t: Vec<String> = vec!["delta-v1".to_string(), self.turn_advance.to_string()];
+        t.push(self.domain_deltas.len().to_string());
+        for d in &self.domain_deltas {
+            t.push(d.domain.0.to_string());
+            for &v in &d.observed {
+                t.push(v.to_string());
+            }
+            for &v in &d.hidden {
+                t.push(v.to_string());
+            }
+        }
+        t.join(" ")
+    }
+
+    /// Reconstruct a delta from [`serialize`]. Total: `None` on malformed input.
+    pub fn deserialize(s: &str) -> Option<WorldDelta> {
+        let mut it = s.split_whitespace();
+        if it.next()? != "delta-v1" {
+            return None;
+        }
+        let turn_advance: u64 = it.next()?.parse().ok()?;
+        let n: usize = it.next()?.parse().ok()?;
+        let mut domain_deltas = Vec::with_capacity(n);
+        for _ in 0..n {
+            let domain = DomainId(it.next()?.parse().ok()?);
+            let mut observed = [0i64; LANES];
+            for v in observed.iter_mut() {
+                *v = it.next()?.parse().ok()?;
+            }
+            let mut hidden = [0i64; HIDDEN_LANES];
+            for v in hidden.iter_mut() {
+                *v = it.next()?.parse().ok()?;
+            }
+            domain_deltas.push(DomainDelta { domain, observed, hidden });
+        }
+        Some(WorldDelta { domain_deltas, turn_advance })
+    }
 }

plan.md

@@ -12,6 +12,54 @@ single executor behavior
 single rune behavior
 single hidden damage formula
 decorative world state
+Compliance Model Anti-Collapse
+The framework must also resist collapse of the compliance model.
+
+The implementor may not replace a specification obligation with a representative approximation that merely resembles the obligation.
+
+Every acceptance requirement must have all of:
+
+1. A measured artifact.
+2. A provenance chain from the measured artifact to the run that produced it.
+3. A merge-blocking enforcement point.
+4. A failure condition that blocks acceptance if the artifact or provenance is absent.
+
+No requirement may be satisfied by:
+
+sample
+summary
+approximation
+representative subset
+default profile
+proxy metric
+regenerated artifact
+local-only executable
+documentation claim
+
+unless this specification explicitly defines that weaker substitute as acceptable for that requirement.
+
+The compliance path must be:
+
+Specification requirement
+Mandatory enforcement mechanism
+Merge blocked if absent
+
+The compliance path must not be:
+
+Specification requirement
+Representative approximation
+Evidence of approximation
+
+Examples of forbidden substitutions:
+
+100% reference/runtime comparison may not be replaced by comparison of base executions only.
+Required artifacts may not be replaced by partial artifacts.
+Full trace information may not be replaced by a summarized proxy unless that proxy is explicitly named as the acceptance artifact.
+Persisted expectations may not be replaced by expectations regenerated in the same run.
+Merge-blocking enforcement may not be replaced by a manually runnable local binary.
+
+If an implementation uses a weaker substitute, the correct result is not partial credit. The correct result is failure of the corresponding acceptance gate.
+
 Hard CI Gates
 Minimum per full CI run:
 

plan2.md

@@ -0,0 +1,262 @@
+
+Web Game Implementation Plan
+Prime Directive
+Build the browser game around the already-defined Rust simulation/testing system.
+
+The web game must be:
+
+Browser-first
+Server-authoritative
+Replayable
+Deterministic
+Test-gated
+Playable before pretty
+No web feature may bypass the Rust runtime contract.
+
+Architecture
+Rust simulation core
+        ↓
+Authoritative server
+        ↓
+WebSocket protocol
+        ↓
+Browser client
+        ↓
+UI / arena / rune editor
+Browser never decides truth.
+
+Browser only sends:
+
+intent
+movement choice
+rune program
+library slot selection
+inspection request
+Server returns:
+
+world snapshot
+visible observations
+turn result
+trace excerpts
+replay hash
+legal player-facing diagnostics
+Crates
+crates/
+  world_model
+  rune_ir
+  trace_model
+  reference_runtime
+  game_runtime
+  replay_corpus
+  protocol
+  server
+  web_client
+  web_assets
+  web_tests
+Phase A — Protocol First
+Define all client/server messages before UI.
+
+ClientMessage
+  JoinMatch
+  SubmitTurn
+  EditRuneProgram
+  InspectTarget
+  RequestReplay
+  Ping
+
+ServerMessage
+  MatchState
+  TurnStarted
+  TurnResolved
+  ObservationResult
+  ValidationReport
+  ReplayChunk
+  ErrorEvent
+Rules:
+
+All messages versioned
+All messages serializable
+All messages replay-testable
+All server outputs hashable
+No client-only game truth
+Phase B — Authoritative Match Server
+Server owns:
+
+match state
+turn timer
+submitted actions
+rune execution
+visibility filtering
+knowledge filtering
+replay recording
+disconnect handling
+Server loop:
+
+Create match
+Send visible snapshot
+Start turn timer
+Collect actions
+Resolve through runtime
+Persist replay event
+Send filtered results
+Advance turn
+Hard gates:
+
+same inputs produce same replay hash
+late input rejected deterministically
+disconnect does not corrupt match
+invalid client packet cannot panic server
+client cannot mutate hidden state
+Phase C — Browser Client Shell
+Client responsibilities:
+
+connect
+authenticate anonymously/dev
+join match
+render visible arena
+show entities
+show turn timer
+edit rune program
+submit action
+display results
+display observations
+play replay events
+Do not implement complex art yet.
+
+Use debug visuals:
+
+grid
+tokens
+panels
+logs
+timers
+entity markers
+domain indicators
+Phase D — Rune Editor
+The rune editor is the core UI.
+
+Required:
+
+keyboard-bound rune input
+token grid / sequence view
+library slot panel
+syntax-neutral execution preview
+visible cost/risk diagnostics
+submission lock on timer expiry
+Important:
+
+The editor must not pretend to know full truth.
+It can show observed diagnostics only.
+Player-facing diagnostics should say:
+
+known reads
+known writes
+observed risks
+unknown listeners
+previous outcomes
+Not:
+
+guaranteed damage
+guaranteed success
+full hidden state
+Phase E — Arena Interaction
+Each turn, player can:
+
+move
+inspect
+cast rune program
+use stick/basic attack
+wait
+All actions become server commands.
+
+Client-side previews are advisory only.
+
+Phase F — Visibility / Knowledge Layer
+Server sends filtered state:
+
+VisibleWorldSnapshot {
+  observed_domains,
+  observed_entities,
+  observed_environment,
+  known_history,
+  inferred_markers,
+  hidden_state_redactions,
+}
+Knowledge must be game state, not UI notes.
+
+Client displays:
+
+known
+unknown
+suspected
+contradicted
+newly observed
+Phase G — Replay System
+Every match produces:
+
+initial seed
+player inputs
+turn boundaries
+runtime hashes
+visible outputs
+trace excerpts
+final hash
+Browser replay consumes the same protocol stream.
+
+CI gate:
+
+recorded replay equals regenerated replay
+browser replay event order matches server order
+Phase H — Web Testing
+Required test layers:
+
+Rust protocol tests
+server integration tests
+browser protocol tests
+Playwright end-to-end tests
+replay determinism tests
+fuzzed packet tests
+disconnect/reconnect tests
+timer edge tests
+Minimum web CI gates:
+
+1,000 simulated matches
+10,000 protocol fuzz cases
+100 browser E2E matches
+0 server panics
+0 replay hash mismatches
+0 hidden-state leaks
+Phase I — Vertical Slice
+First playable slice:
+
+2 players or 1 player + dummy opponent
+small arena
+turn timer
+movement
+inspection
+basic attack
+rune submission
+multicast execution
+visible consequences
+replay viewer
+No progression.
+No accounts.
+No cosmetics.
+No marketplace.
+No complex content.
+
+Acceptance Criteria
+Web phase is accepted only when:
+
+A player can join a browser match.
+A turn timer runs.
+The player can inspect, move, attack, or cast.
+Rune programs execute only on the server.
+Results return as filtered observations.
+Replay can reproduce the match.
+Browser cannot alter hidden truth.
+CI proves protocol, replay, visibility, and server authority.
+Core rule:
+
+The web game is just a playable window into the Rust universe.
+It must not become a second simulation.