test: weekly mutation-score ratchet over decision kernels (Stryker, non-gating first)

[thymikee]

Part of #1412 (Track A).

Context

When agents write both code and tests, mutation score is the only mechanical answer to "is this test load-bearing or decorative". Full-suite mutation is off the table; this targets an enumerated list of pure decision kernels where a surviving mutant means a silently wrong agent-facing decision.

Deliverables

• Stryker (vitest runner) config scoped to exactly: src/kernel/errors.ts (retriability/hints), src/daemon/ref-frame.ts (ADR 0014 admission matrix), src/commands/interaction/runtime/settle.ts, src/utils/scroll-edge-state.ts, selector matching (src/selectors/).
• Weekly scheduled workflow; report as job summary + artifact. Non-gating initially.
• A per-module score baseline file with a ratchet check (scores may only rise), enabled once two consecutive weekly runs are stable.

Acceptance

• Weekly run completes < 30 min on ubuntu runner.
• Baseline file + ratchet script tested (a lowered score fails the check with the surviving-mutant list).

Out of scope

Any module that spawns subprocesses or waits real time; gating PRs on mutation score.

Review amendments (2026-07-27)

• Define graduation criteria up front: after N stable weekly runs, mutation testing gates PRs for AFFECTED modules only (changed file in the kernel list → run its mutants on the PR); weekly stays the full sweep. Otherwise regressions are found only post-merge.
• Store Stryker version + config content hash alongside each baseline; a score change caused by a tool/config change must be distinguishable from a test-strength change.

PR-wave review outcome (2026-08-01, PR #1441)

Approve-with-changes; real StrykerJS, exact five-kernel scope asserted at three points, both amendments implemented precisely (provenance-tagged baselines, drift vs regression classified, graduation mechanically enforced). Remaining items now part of this spec:

• Strip or split the bundled obs: scheduled-lane health, freshness telemetry, and standard artifact envelope #1430 watcher; KEEP scripts/lib/lane-envelope.ts as the wave's shared envelope module.
• Budget verification before trusting the schedule: kernel-errors measured 9m04s/183 mutants on a 2-core runner → the 1,277-mutant selectors shard projects to ~45–55 min, past the <30-min criterion and the 45-min shard timeout. One manual workflow_dispatch is required; re-shard or trim globs if the projection holds.
• Maintainer decision required: mutation-affected.yml (PR-triggered, dormant-by-baseline) already spends real CI minutes pre-graduation. Approve the spend explicitly or gate the trigger on stableRuns >= 2.
• Document the intentionally-missing tsconfig.stryker-absent.json trick with a link to the Stryker behavior it relies on.

[thymikee]

Scope amendment (recorded per the PR #1553 review): the five-kernel Stryker boundary approved here is extended by two modules — packages/ad-script/src/internal/target-annotation-serde.ts and src/snapshot/snapshot-occlusion.ts. Both satisfy this issue's own membership rule (pure decision kernels, zero I/O, where a surviving mutant is a silently wrong agent-facing decision); the gap was found by the 2026-08-01 post-#1478-arc test-strength audit, and the maintainer initiated the implementation task from that audit. packages/replay-test was evaluated and excluded (spawns subprocesses, fails the rule). Implementation: PR #1553.

Generated by Claude Code