feat: replayable C0.4 golden corpus (registry v2)

[Freshair129]

Summary

This PR closes GKS-MIG-002, one of the owner-decision rows in the P1 report. The owner chose to ship the C0.4 golden corpus as something that can actually be replayed. Design and evidence: docs/reports/2026-09-27-c0-corpus-v2.md.

Before (v1): each case had a request hash and an expected-result hash, but not the request or the result. Nothing could recompute the hashes, and every PASS pointed at ordinary contract tests.

Now (v2): each runnable case ships three things:

• the requests to send, in cases/<id>.json;
• the normalized answer to every step, in expected/<id>.json;
• a runner that replays the case against a real gks-server stdio process on a fresh SQLite store.

What a replay checks

A case passes only when all five hold:

1. the request fixture still hashes to the registry;
2. every step's intent annotation holds (ok, a named tool error, a protocol error message, or a partial match);
3. stdout carries JSON-RPC frames only, and no frame arrives unrequested;
4. no fixture credential appears in stdout, stderr, or the SQLite main, WAL or SHM files;
5. the normalized transcript hashes to the registry.

Normalization. Three kinds of value change from run to run and are replaced by stable labels before hashing:

• server-clock instants: an instant the request itself carried is left as is;
• measured durations;
• hashes derived from the clock, such as decisionHash and the receipt hashes chained from it. Equal hashes keep equal labels, so an idempotent replay stays visible in the transcript.

A request that needs one of those hashes takes it from an earlier response with $bind.

Re-baselining. --write records a case only when two consecutive runs agree. A contract test rebuilds the fixtures and fails if they differ from the committed files, so a change to a shared helper cannot silently change the corpus.

Cases (25; 24 runnable)

• 17 tool cases: each calls its own tool. initialize and tools/list pin the protocol surface.
• API-010 replay: replay across a restart, plus a conflict for a changed payload.
• Tenant wall.
• Auth denial (secure mode): missing auth, forged credential, wrong role, stale version, forged scope digest.
• Transport denial: malformed JSON, duplicate keys, a batch, deep nesting, oversized frame, invalid UTF-8, non-finite number.
• GenesisRAG17 receipts:
  • the full chain through publication;
  • wrong hash, scope, stage or role is denied;
  • a batch whose Stage 15 fails gets a gate verdict FAIL with allowPublication: false.
• Backend failure: a table is dropped from a second connection, which gives gks_backend_unavailable and no partial write.
• Lost-response replay: SIGKILL after the commit, for both a promote and a submit.
• C0.4-TIER4-READBACK (NOT_RUN): split out of the v1 receipts case. GenesisBlockDB's physical readback cannot run inside GKS, which never calls outward.

The C0.4 result manifest now records PASS 24 and NOT_RUN 1. productionReady and deploymentAuthorized stay false. The real-MSP runs recorded at be97c93 move to externalRuns with their original date.

Gating

• check:c0 now also checks, without running anything, that each fixture and transcript hashes to the registry, and it allows a PASS only for a case that can be replayed.
• New npm run check:corpus. The c0-gate CI slice runs it after check:c0 and check:baseline.
• tests/integration/c0-corpus.test.mjs runs the same replay inside npm test.
• The baseline lock is re-locked because the workflow hash changed.

Behaviour

This PR does not change any product code. It only records how the server behaves today.

One thing I noticed is left unchanged: when a batch has no worker receipt, the gate gives the security reason "retrieval benchmark reported a cross-tenant leak", which is misleading because no benchmark exists at all. It is noted in the report for a later PR.

Test plan

• npm run check:corpus: all 24 cases replay, in about 3 s, identical across 5 consecutive runs. --write also ran each case twice.

• npm test: vitest 288 passed and 2 skipped (the MSP suites need MSP_REPO_ROOT). Security 12/12, unit 9/9.

• npm run check:c0 gives PASS 24 / NOT_RUN 1, and npm run check:baseline holds.

• Mutation checks: each mutation was applied to product code, the corpus was run, and the mutation was reverted.

  Mutation	Failing cases
  parse-error message	TRANSPORT-DENIAL
  idempotent flag on replay	TOOL-012, GENESISRAG17-RECEIPTS
  aliases dropped	4 cases, caught by the golden hash alone
  normVersion value	4 cases, caught by the golden hash alone
  server fails to start	all 24, each at once with the server's stderr

• RKOI architecture review. It was still running when this PR was opened; any findings will be fixed in follow-up commits on this PR.

• CI on this PR.

🤖 Generated with Claude Code