feat(repo): agent-ready failure diagnostics (stack #146-#156)

[ryanleecode]

Summary

Lands the agent-ready failure diagnostics stack (plan docs/plans/2026-10-01-2306-feat-agent-ready-failure-diagnostics-plan.md) on main. The eight stacked PRs were merged into this feature branch; this PR brings the branch into main with a merge commit.

• fix(repo): draw every deciding timeout class in the timeout-fields property #146 draw every deciding timeout class in the timeout-fields property
• feat(repo): end every failed run in one typed failure record #147 end every failed run in one typed failure record
• test(repo): declare the deciding coverage classes of two under-drawn properties #148 declare the deciding coverage classes of two under-drawn properties
• feat(repo): give every failure record a one-step reproduction #149 give every failure record a one-step reproduction
• feat(repo): carry the failure record to MCP, MSP, SARIF and traces #150 carry the failure record to MCP, MSP, SARIF and traces
• ci(repo): render mutation job failures from the failure record #151 render mutation job failures from the failure record
• ci(repo): dry-run the main mutation lane once and keep its evidence #155 dry-run the main mutation lane once and keep its evidence
• ci(repo): run the mutation lane on pull requests and gate new survivors #156 run the mutation lane on pull requests and gate new survivors

Merge from main

main gained #153 and #154 after the stack was cut, so the branch carries one merge commit from main with these conflicts resolved:

• worker-client.blueprint.ts: main's boot timeout (timeoutOrElse with WORKER_BOOT_TIMEOUT, no connect retry), plus the stack's workerKind on WorkerBootTimeoutError.
• substituted-worker.fixture.ts: bootPingWorker, which fix(repo): fail the run instead of hanging when a plugin worker never accepts its socket #154 moved here, now passes workerKind: 'testRunner'.
• dry-run-failure.integration.test.ts: main's runWithRunnerPlugin and deaf-runner scenario, written against the stack's RunFailure. The stage is read from evidence.stage, and there is no log capture because the stack asserts the failure record instead of logs.
• worker-launcher.integration.test.ts: main's version, since the helpers moved into the fixture.

Testing

• Local, on the merged tree: build and typecheck of @systemfsoftware/stryker-js and its dependencies pass. gate:tasks passes 111 of 112 tasks. The one failure was git-diff.integration.test.ts, which failed only because my local global git config signs commits with gpg; with a clean HOME it passes 2/2. The conflict-area suites (dry-run-failure, worker-launcher, worker-boot-timeout, clean-temp-dir) pass 14/14.
• Real CLI run on test/e2e/testResources/failing-fixture:
  • exit code 5;
  • the BaselineTestsFailed record names src/thing.test.ts:7:21, with next action fixCode, otherwise fixTest;
  • reports/mutation/failure.json is written;
  • the printed reproduce: command reruns exactly the failing test.
• CI on the stack PRs:
  • ubuntu check and every e2e lane are green on each head after reruns. The rest-1 LeakedState failure on feat(repo): give every failure record a one-step reproduction #149 was a flake.
  • macOS timed out in vm-parity on ci(repo): render mutation job failures from the failure record #151 three times. The same code is green on macOS in ci(repo): dry-run the main mutation lane once and keep its evidence #155 and ci(repo): run the mutation lane on pull requests and gate new survivors #156.
  • On ci(repo): run the mutation lane on pull requests and gate new survivors #156, the stryker-js mutation shard hits the 1800 s cap. As ci(repo): run the mutation lane on pull requests and gate new survivors #156 explains, PR shard plans keep reading main's old-shape timings, which underestimate a full-cost run, until main records its first timing with the new fields.

Follow-ups

• mutation.yml uses cancel-in-progress: true keyed on github.ref, which was already true on main. A second merge to main cancels the full main run that the PR lane now depends on for its survivor baseline and timings. Main should probably queue rather than cancel.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}