[dev-pmallapp]

Build target 6 of docs/design/stories/293.md. Anchors CLAIM-293.3 and NEVER-293.4.

Verdict: PASS for cases 3, 5, 6, 7, 11, 21 and 22. CLAIM-293.3 and NEVER-293.4 are met for the harness's transcription of each skill — a different population from the skill, and the runner says so in its own output on every run.

Evidence: docs/evidence/322-20260910T042657Z.md

bun run skill-harness exists for the first time. The command named in seven test-plan cases was in no package.json in this tree until now.

Chain

1038 tests / 1024 pass / 0 fail / 14 skip / 3777 expect() · lint 0 across 121 files · typecheck clean · skill-lint 4 + 12, 0 errors · claim-lint 120 files, 0 errors, 1 warning (the expected stories/9.md one) · bun run skill-harness exit 0 over six scenarios · both verifier scripts PASS with the required contexts still six — the seam rung's CI job is #323's six-site edit and is deliberately not here.

What it ships

• packages/harness/src/runner.ts — runHarness, and decideVerdict, the one pure function that turns typed observations into an exit code. Thirteen denominators, each reported on its own line whether it passed or failed.
• Four transcriptions, transcription-<verb>.ts, one step per ## Re-entry row, each read attributed to the row it decides.
• scenario-roster.ts — six frozen scenarios covering 4 of 4 skills, including a forge-only and a worktree-only scenario so neither surface can be silently unwired.
• re-entry.ts and argv-kind.ts — the Re-entry parser promoted out of a test file so there is exactly one, and a positive argv classifier so a read is never computed by subtraction.
• scripts/skill-harness.ts — the only writer; it passes verdict.exitCode to process.exit with no branch.

Measured, per skill

goal-create 3 Re-entry rows / run 2 read 3 · story-create 4 / 4 · story-design 4 / 4 · story-test-plan 5 / 5 — every threshold read off skills/*/SKILL.md at run time. That is what separates "decided to do nothing" from "did nothing".

Mutation testing — 16 injected, 13 killed, 3 survivors

Three mutations survived their first run. Each was a real gap, each is now closed and re-killed — that is the most useful part of this PR:

• M10 crashed instead of reporting: a fresh forge before run 2 made a transcription throw, and the runner died with no verdict, no artifact, no scenario named. Now a scenario-threw failure, one per cause.
• M11 was behaviourally equivalent — deriving the exit code from renderReport(...).includes("verdict: pass") gives the identical answer for every fixture, so no dynamic test could ever kill it. Killed by a static check on decideVerdict's own source. Finding 37, whose recorded owner was any future harness.
• M12 exposed an unasserted field: renderArtifact could emit verdict: "pass" on a failing run and everything passed. S2.6.6 — The seam rung and its CI job #323's CI job may read that file, so the disagreement would have surfaced as a green job over a red run.

M4, M5 and M7 leave bun run skill-harness exit 0 and unchanged. They die only to synthetic scenarios built to trip them:

A green run of the real roster is not evidence that the runner's checks work. It is evidence that the roster passes them.

M9 survived everywhere and is kept — run2.treeHash == run1.treeHash is entailed by run 2's evidence being empty over this roster; it is retained because the issue mandates it and M10 is its only independent killer. S1 survives by design: reordering a ## Re-entry row changes nothing, which is NEVER-293.5 working.

Findings

1. S2.6.4 — The two-surface mutation recorder #321's write-seam lint was red against correct code. It matched the bare word writeFile, and a context method that only delegates to the recorder tripped it. The repository had already recorded this class — port.test.ts rewrote its launcher detector from a call shape to an import for the same reason — and it was reproduced one task later anyway. The detector now reads the capability off the import list, with a regression test.
2. A read may not be computed by subtraction. isMutatingGhArgv is false for every git argv, so calls.length - attemptedMutations.length scores git commit as a read.
3. The fake models neither a milestone PATCH nor --milestone on gh issue edit. Two Re-entry branches are consequently dead here — for a correct reason — and were left unexercised rather than forced around the fake's gaps.

Not established, deliberately: the transcriptions are unaudited (CLAIM-293.11, case 18 — #323); nothing tests the model reading a SKILL.md; CLAIM-293.6 half (ii) is still NOT MET; no CI job; #330 is still unpaid — every gh call was answered from #320's fake; #331 untouched.

Refs #322