[dev-pmallapp]

Refs #315. Refs #47. Build targets 1 and 2. Anchors CLAIM-47.7, CLAIM-47.1. Cases 2 and 12.

No Closes directive, deliberately. NEVER-47.8 — this Story's own anti-claim — forbids a task PR carrying one, and CONTRIBUTING.md:415 records that it would never fire against a story branch anyway.

The defect, measured before anything was changed

At a6ce084, packages/core/src/gh/sub-issues.ts:296-304 built gh issue edit <parent> --repo … --body <checklist> where the body was tasksChecklist's output — only ## Tasks\n\n- [ ] #905 …. The Story's entire body was replaced by its own checklist.

--- emitted `gh issue edit` --body, in full ---
"## Tasks\n\n- [ ] #905\n- [ ] #906"
--- does it retain ANY of the Story body? ---
  ## Problem           retained: false
  ## Claims            retained: false
  CLAIM-99.1           retained: false
  a reviewer reads     retained: false

packages/core/test/gh-sub-issues.test.ts:165-169 asserted the first four argv elements and that the last contained ## Tasks. Both are true of a body that has destroyed everything else. A passing test that is wrong.

What shipped

withTasksChecklist(body, items) — build target 1, modelled on withParentLine one file over: it inserts into an existing body, it is idempotent, and it refuses the ambiguous case rather than guessing.

Behaviour	Why
Rewrites only the leading run of entry lines, never the whole section	Real Story bodies put load-bearing prose inside ## Tasks — see below
Refuses a body with two ## Tasks sections, naming the count	Two checklists are two answers to "what are this Story's tasks"
Refuses entries split by prose, naming the line	The structure is asserted, not assumed — #323's lesson
Inherits a tick the caller is silent about; an explicit checked wins	A tick is body content too
Rebuilds items through safeOwnValue, never a spread	A spread runs a hostile getter; this module requires 0 throws

parentBody is required on both paths and has no default. A default of "" would make the destructive behaviour the fallback for a caller who simply forgot — the defect with a nicer signature. A genuinely empty Story body is stated as "".

Build target 2: the checklist edit is emitted on the GraphQL path too. At a6ce084 that path returned exactly one command and no checklist, so CLAIM-47.1's clause "adds a ## Tasks checklist to the Story" was unreachable on the primary path. Both paths now build the edit through one helper, so they cannot drift on what "adds a checklist" means.

The finding, and it came from the one survivor

M11 — widening the section boundary from ^#{1,2} to ^#{1,3} — survived the first pass, changing not one assertion. Chasing it down found the real defect:

Every real Story body carries load-bearing prose INSIDE the ## Tasks section, after the entries. #47 carries its dependency ordering and the "must not be built twice" ruling there. #293 carries its sequencing rulings there — and on #293 that section runs to the end of the body, so no following heading would stop a section-wide rewrite.

The merger as first written replaced the section from its heading to the next H1/H2. Against #293 it would have deleted every one of those rulings. The fix for a body-destroying defect had reproduced the defect one level down, inside itself, with the full suite green across it.

A green suite over the corpus you thought of is not evidence about the corpus that exists. Every fixture had prose around the checklist; not one had prose inside it, which is what every real Story has.

Mutations — 13 run, 11 killed, 2 void

#	Mutation	Verdict
M1	Case 12. The merge becomes an overwrite — the a6ce084 defect restored	KILLED
M2	The append path drops the body it was appending to	KILLED
M3	The rewrite drops everything after the section	VOID — pattern gone; M13 carries its intent
M4	Every listed box treated as ticked	KILLED
M5	Tick inheritance removed: a human's ticks silently cleared	KILLED
M6	Two ## Tasks sections stop being refused	KILLED
M7	The section's trailing blanks swallowed	VOID — pattern gone
M8	parentBody silently defaults to ""	KILLED
M9	Build target 2 undone	KILLED
M10	safeOwnValue replaced by direct reads	KILLED
M11	An H3 inside the section becomes a boundary	SURVIVED, then KILLED
M12	The structure assertion made vacuous	KILLED
M13	The rewrite swallows the whole section, prose and all	KILLED

M3 and M7 are VOID, not survivors — their patterns stopped existing when the code they targeted was replaced. A mutation whose pattern does not match has measured nothing.

Case 12 is a forward-dated kill and the pre-landing state is recorded as a measured gap, never as a pass. At a6ce084 there was no preservation logic to delete — grep -c withTasksChecklist over packages/core/src returns 0 — so the mutation was inapplicable there, not surviving. M1 pays the kill now.

Two process findings recorded in the artifact

A revert is a mutating operation. The mutation script's git checkout -- <path> ran against a dirty tree and destroyed the section fix before it was committed. Its own post-revert byte-identity assertion fired on mutation 1 rather than at the end of a 13-mutation round; it now refuses to start while those files are dirty.

The task-branch recipe pushes to the story branch. git checkout -b task/… origin/story/… sets the upstream to the story branch, and with push.default = tracking a plain git push -u origin task/… lands three commits directly on it. It did. story/47-s2-3-execution-skills was restored to a6ce084 and the commits re-pushed to their own ref. The existing task/196-revert-direct-push branch suggests this is the second occurrence.

What this does not close

• Case 1 is not claimed. Its ## Build Targets parser is build target 3, owned by S2.3.1 — task-create #48. Only the checklist half of CLAIM-47.1 is touched here.
• No gh argv this repository builds has ever been executed, and no build target owns paying that bill #330 — no gh argv built here has been executed. Both paths are verified as argv only. A seventh task defers it.
• planSubIssueAttach has no production caller yet (task-create is S2.3.1 — task-create #48), so this contract change is free today and will not be later.

Chain, measured at db04726

bun test        917 tests / 903 pass / 0 fail / 14 skip / 3241 expect()
bun run lint    0 violations across 95 files
bun run typecheck   clean
bun run skill-lint  4 SKILL.md + 12 contract files, 0 errors, 0 warnings
bun run claim-lint  114 files, 0 errors, 1 warning
IAI_PURITY_CHILD=1        9 pass
IAI_BINDING_SWEEP_CHILD=1 5 pass
scripts/verify-required-checks.sh   PASS, six contexts
scripts/verify-workflow-hygiene.sh  PASS

Baseline at a6ce084: 903 tests / 889 pass / 3179 expect(). The expect() total becomes 3243 once the evidence artifact lands — #299, and the delta is a property of the corpus an artifact joins, verified rather than assumed.

Evidence: docs/evidence/315-20260920T112814Z.md