[dev-pmallapp]

Build target 5 of docs/design/stories/293.md. Anchors CLAIM-293.8.

Verdict: PASS for case 4 and case 10. CLAIM-293.8 is met for the recorder; its per-skill threshold — at least one scenario per file-writing skill with run1.tree_hash != pre.tree_hash — needs a transcription and a runner and closes at #323, which must not read this as having closed it.

Evidence: docs/evidence/321-20260909T125257Z.md

Chain at 761bb1d

987 tests / 973 pass / 0 fail / 14 skip / 3513 expect() · lint 0 across 109 files · typecheck clean · skill-lint 4 + 12, 0 errors · claim-lint 117 files, 0 errors, 1 warning (the expected stories/9.md one). The evidence artifact adds +2 expect() on top, per #299.

What it ships

• packages/harness/src/mutation-recorder.ts — two surfaces, reported as two separately named lists, never one total. The forge surface reuses isMutatingGhArgv; there is no second argv classifier.
• A write seam, because the third clause is an ATTEMPT clause. git status, git hash-object, a --stat diff and an mtime comparison all sample effect, and the defining property of a byte-identical rewrite is that it leaves no effect. The seam is the worktree's createRecordingPort.
• Three worktree observables, not one. git commit --allow-empty after a byte-identical rewrite moves HEAD while leaving HEAD^{tree} byte-identical, so a tree-hash-only recorder cannot tell a committed identical-bytes rewrite from an uncommitted one.
• No verdict anywhere. mutationEvidence returns a list whose emptiness is the zero-mutation result, mirroring ghOk([]). A boolean is the assertion shape under which every surface-unwiring mutation survives.
• attemptedMutations is typed GhPlan — an argv list with no exit code — so && c.exitCode === 0 is a type error downstream rather than a review miss.

Mutation testing — 14 injected, 12 killed, 2 survivors by design

M3 and M4 kill disjoint halves: if (identical) return; kills only the two identical-bytes scenarios, and identical := existedBefore kills only the differing-bytes ones. A suite without case 10 ships M3 green.

M5/M6 is #320's M9/M9b one layer up: weakening the matrix assertion to a length check survives alone, and the kill for the paired defect comes from a different test. A weakened assertion is not detectable by the test it weakens.

M12's asymmetry: an exit-code filter is killed only by a lie that refuses. The silent-no-op lie exits 0, so the property "records the attempt, not the success" is observable only through a double that says no.

S2 survives — the fake made to enforce a duplicate leaves this suite green — while M12 dies. Same property, opposite sides.

Findings

1. Every effect-sampling alternative to the write seam is blind to an identical-bytes rewrite by construction.
2. The seam has a run-time bypass detector, and its blind spot is the same one the seam exists to cover — stated in the module, the test and the artifact so nobody reads it as complete.
3. Tracked-ness is recorded, never filtered. A Design is untracked at the moment it is created; filtering on it would score story-design as zero mutations.
4. M8's prediction was wrong and is recorded as wrong — it kills C as well as E. The conclusion held; the stated mechanism did not.

The first-contact bill (#330) is still unpaid — this task answered every gh call from #320's fake. #331 is untouched: the transient mode was injected to prove a failed attempt is recorded, and no retry logic was built on it.

Refs #321