Filed unassigned by the dev seat on #8246 (PR #8365) — out of that card's graded lane, recorded rather than fixed in flight.
The shape
#8246 established that a dogfood ablation run against a stale dist/ goes green and silently certifies a vacuous test, and that no later CI run can expose it. Its graded scope put the remedy in two agent-facing documents (.claude/agents/os-dev.md, .claude/skills/dogfood-verification/SKILL.md) plus a pre-flight script, all landed in PR #8365.
There is a third live copy of the same instruction, and it is the one physically closest to the suite — packages/qa/dogfood/README.md, section "Adding a golden test", step 4:
- Prove it catches the bug: temporarily revert the relevant fix and confirm the test goes red. A green-on-the-bug test is not a gate.
That is the ablation procedure, stated correctly as far as it goes, and it stops exactly where the incident begins: it does not say to rebuild the package between "revert the fix" and "confirm the test goes red", and the reader of this file is by definition someone adding a test to the suite that consumes dist/. The false green is available to anyone who follows the README instead of the agent definition.
Why it was not folded into #8246
The grading comment on #8246 (2026-08-12 23:50Z) fixed that card's scope to the agent-definition/skill lane, and its triage comment warned that a packages/qa landing would re-route the card to another lane (domain:cli per the package table). Editing this README is a small change but it belongs to that other lane, so it is filed here rather than smuggled into a domain:skills PR.
Suggested fix (small)
One clause in step 4, pointing at the tooling PR #8365 lands:
- rebuild the package after reverting, before judging the colour;
node scripts/ablation-dist-preflight.mjs PKG 'MARKER' (or --absent for a deleted guard) exits non-zero unless the consumed dist/ really carries the mutation.
Keeping the three copies structurally consistent is the point — the README copy is the one a human contributor is most likely to read first.
Refs: #8246, PR #8365.
Blocked-by: #8246
(Triage 2026-08-13: the step-4 clause this card prescribes cites scripts/ablation-dist-preflight.mjs, which lands only in PR #8365 — an ADR-class draft awaiting the maintainer's hand-merge. Dispatching before #8246 closes would document a script that is not on main.)
Filed unassigned by the dev seat on #8246 (PR #8365) — out of that card's graded lane, recorded rather than fixed in flight.
The shape
#8246 established that a dogfood ablation run against a stale
dist/goes green and silently certifies a vacuous test, and that no later CI run can expose it. Its graded scope put the remedy in two agent-facing documents (.claude/agents/os-dev.md,.claude/skills/dogfood-verification/SKILL.md) plus a pre-flight script, all landed in PR #8365.There is a third live copy of the same instruction, and it is the one physically closest to the suite —
packages/qa/dogfood/README.md, section "Adding a golden test", step 4:That is the ablation procedure, stated correctly as far as it goes, and it stops exactly where the incident begins: it does not say to rebuild the package between "revert the fix" and "confirm the test goes red", and the reader of this file is by definition someone adding a test to the suite that consumes
dist/. The false green is available to anyone who follows the README instead of the agent definition.Why it was not folded into #8246
The grading comment on #8246 (2026-08-12 23:50Z) fixed that card's scope to the agent-definition/skill lane, and its triage comment warned that a
packages/qalanding would re-route the card to another lane (domain:cliper the package table). Editing this README is a small change but it belongs to that other lane, so it is filed here rather than smuggled into adomain:skillsPR.Suggested fix (small)
One clause in step 4, pointing at the tooling PR #8365 lands:
node scripts/ablation-dist-preflight.mjs PKG 'MARKER'(or--absentfor a deleted guard) exits non-zero unless the consumeddist/really carries the mutation.Keeping the three copies structurally consistent is the point — the README copy is the one a human contributor is most likely to read first.
Refs: #8246, PR #8365.
Blocked-by: #8246
(Triage 2026-08-13: the step-4 clause this card prescribes cites
scripts/ablation-dist-preflight.mjs, which lands only in PR #8365 — an ADR-class draft awaiting the maintainer's hand-merge. Dispatching before #8246 closes would document a script that is not onmain.)