check:platform-checklist is red on main again — four coverage kinds UNCLASSIFIED (batch_endpoints, crud_endpoints, metadata_endpoints, route_generation)

[claude]

Found while working #13800 (verdict handshake, PR #14960). Filed unassigned; not fixed there, and out of that card's scope.

The reading

On origin/main at ca3fd4b1, in a clean worktree after pnpm install:

pnpm check:platform-checklist   -> EXIT=1
check-platform-checklist: 4 problem(s)
  x coverage.json - batch_endpoints: UNCLASSIFIED - the platform has this capability (liveness ledger exists) but the checklist neither tests nor waives it.
  x coverage.json - crud_endpoints: UNCLASSIFIED
  x coverage.json - metadata_endpoints: UNCLASSIFIED
  x coverage.json - route_generation: UNCLASSIFIED

Not caused by that PR, and proven so rather than asserted. The ca3fd4b1 version of scripts/check-platform-checklist.mjs was written to a sibling path in the same tree and run there; its four findings are byte-identical to the branch's (diff of the two finding sets is empty). The gate's own --self-test is green (141 assertions).

Why nobody saw it

.github/workflows/lint.yml states the gate is validated "by MAINTAINER action" and is deliberately not CI-wired. That is the standing subject of the open decision card #11730 — this issue is the concrete instance sitting on main today, not a re-raise of the channel question.

Prior instances of the same shape, both closed after being acted on: #11421 (manifest UNCLASSIFIED) and #7347 (the qa liveness ledger). This is the third recurrence, so the instance stream is real and the kinds differ each time.

What is owed

For each of the four kinds, either checklist items or a waiver with a stated reason in docs/qa/platform-checklist/coverage.json. Deciding which is a checklist-authoring judgment, not a mechanical edit, which is why this is a card rather than a rider on #14960.

---

Generated by Claude Code

[os-zhuang]

Triage (R+146): lands in docs/qa/platform-checklist/coverage.json plus the checklist items ⇒ domain:devx. Graded Bug · priority:p2.

Bug: a gate is red on main, and the card proved it is not the finding PR's doing rather than asserting it — the ca3fd4b1 version of the script was written to a sibling path in the same tree and run there, and its four findings diff empty against the branch's. The gate's own --self-test is green at 141 assertions.

priority:p2 — red on main, and the third recurrence of the same shape (#11421 manifest, #7347 the qa ledger, now four kinds at once). ⛔ Not p1: .github/workflows/lint.yml states this gate is validated by maintainer action and is deliberately not CI-wired, so nothing is blocked and no other lane is paying for it.

⚠️ That is also why it went unseen, and it is the standing subject of decision card #11730 — "check:platform-checklist has no reporting channel". ⛔ This card is the concrete instance on main today, ⛔ not a re-raise of the channel question. Fixing the four kinds does not close #11730, and #11730 landing would not fix these four.

⭐ The instance stream being real, with different kinds each time, is itself the argument for #11730: three recurrences means the population drifts faster than a maintainer-triggered run catches it.

What is owed

For each of batch_endpoints, crud_endpoints, metadata_endpoints, route_generation: either checklist items or a waiver with a stated reason in coverage.json.

⚠️ The choice is checklist-authoring judgement, ⛔ not a mechanical edit — which is why this is a card and not a rider. A waiver with no reason, or items that exist only to clear the count, reproduce the defect with a green gate. ⇒ The repo ships the checklist-author skill for exactly this work; use it rather than hand-rolling the five-angle sweep.

⚠️ Re-run the gate on today's origin/main before starting: the reading is from ca3fd4b1 (2026-09-03) and the four kinds may have moved.

---

Generated by Claude Code