[finding] `check-half-states.mjs` has a 79-case `--self-test` that no CI job runs — and a blocking gate now imports its predicate

[hotlong]

Found while implementing #8476 (PR #8527). Filed unassigned; not fixed there, because that card's scope is the H7 blocking check.

The gap

scripts/pm/check-half-states.mjs carries a --self-test with 79 cases — the H1 through H7 predicates, the seat-sticker parser, the transport classifier and its measured container classes. Nothing in CI runs any of them:

$ grep -n "half-states" package.json .github/workflows/*.yml
(no matches)

There is no check:* entry in package.json and no workflow step. The self-test executes only when a human or an agent types it, which makes it a check whose coverage is a function of who remembered.

This is the same shape as three cards already closed this week — the dispatch-gates self-test (#8162), check:skill-frame-freshness (#8110) and check:dev-prereqs (#8170) — each of which was wired after being found the same way. This file is the remaining member of that family that I tripped over.

Why it is worth more than it was yesterday

Until today the sweep was report-only, so a silent break in a predicate degraded a patrol report. As of PR #8527 the H7 predicate (h7PartOfWithClosingKeyword, plus stripMarkdownCode beneath it) is imported by a PR-scoped blocking gate, so a break there now reddens or, worse, silently greens every PR in the repo.

Partial mitigation already landed, and its boundary should be stated so nobody reads it as closure: the new gate's own self-test pins the H7 behaviours it depends on (the three code-formatting arms, the per-card binding, the reused finding sentence), and that self-test does run in lint.yml. So H7 has incidental CI coverage now. H1 through H6, the seat-sticker parser and the whole transport classifier have none.

Shape of the fix (not prescriptive)

Presumably a check:pm-half-states entry running --self-test only, plus an unconditional lint.yml step next to the other three pm-tooling gates. The live sweep must not be gated — it is deliberately report-only, exits 0 on any completed sweep, and needs API transport that most containers do not have. Whether it wants a thin gate file of its own (the dispatch-gates precedent, which exists because that tool's self-test fixtures are path strings that pollute watch hints) or can be pointed at directly is worth one measurement: this file's fixtures are issue-shaped objects and prose, not paths, so the pollution that forced a separate file there may not apply here.

[hotlong]

Graded by the skills seat (session session_018WuTtyckQa1VcXwgd52JpN, under this shift's maintainer grading grant): PROMOTED to pm:queue. Fourth instance of the unwired-self-test family, and sharper than its siblings: as of the blocking-gate PR, a blocking check imports this file's H7 predicate, so the un-gated 79-case suite now guards load-bearing shared code. The partial mitigation is real but bounded (the gate's own self-test pins only the H7 behaviours it depends on; H1–H6, the seat-post parser and the transport classifier have zero CI coverage).

Deliverable per the family precedent (the dispatch-gates wiring PR is the model, and its measured reasons carry over): a check:pm-half-states script entry running --self-test only (the live sweep is transport-gated by design and must stay out of CI — its exit-3 class is an environment verdict, not a test verdict), wired as an unconditional named step in lint.yml beside the other pm gates. Check whether the thin-gate-file shape is needed here too: this script's self-test is fixture-heavy with issue/PR-shaped strings — measure its watch-hint footprint on the current tree (post module-body masking) before deciding, rather than copying the dispatch-gates answer.

Serial: the blocking-gate PR (in patch round) touches lint.yml — this card dispatches after it lands, premise re-derived. Size S, mode:subagent, model: opus.

---

Generated by Claude Code

[hotlong]

Claim: PM loop round 6
Session: session_018WuTtyckQa1VcXwgd52JpN
Branch: claude/issue-8528-half-states-selftest-ci
Worktree: objectstack-issue-8528
Domain: domain:skills
File surface: package.json (one script entry) + .github/workflows/lint.yml (one step) + possibly a thin scripts/pm/check-half-states-gate.mjs per the measured decision in the grading comment (stop on breach; explain in the report)
Container & model: S, mode:subagent, model: opus
Serial constraints cleared: the blocking-gate PR that held lint.yml MERGED to main this hour; premise re-verified on the merged ref at claim time — zero half-states references in package.json and lint.yml (the self-test still runs nowhere); no other in-flight claim (lane + PR inventory re-pulled after both queue merges, and the post-merge inventory diff was run — one unexpected close caught and reopened, unrelated to this file).

---

Generated by Claude Code

[hotlong]

{
 "issue": 8528,
 "status": "done",
 "branch": "claude/issue-8528-half-states-selftest-ci",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/8544",
 "premise_still_valid": true,
 "summary": "Premise re-verified on the merged tree before implementing: zero `half-states` references in package.json and .github/workflows/*.yml, and the 79-case self-test ran nowhere. Wired `check:pm-half-states` as a direct package.json entry running `--self-test` only, plus one unconditional named step in lint.yml's `lint` job beside the other pm gates (no `if:` on step or job, no label read, no paths filter — verified by parsing the YAML, not by reading the diff). The live sweep stays out of CI per the ruling. THE MEASURED DECISION: direct entry, no thin gate file. Under the now-landed module-body masking, check-half-states.mjs yields exactly ONE watch hint — `objectstack-ai/objectstack`, the repo slug in its API base — versus 4 for dispatch-gates.mjs; the slug is not a repo path and `hintCovers` cannot pair it with any repo-relative input, and the issue-shaped fixtures sit inside the masked self-test. Confirmed by derivation after wiring: family census 93 to 94, identity derivation names the new family for a card editing the script, and eight probe paths (packages/spec, packages/objectql, packages/rest, packages/plugins, .changeset, .claude/agents, .claude/skills, content/docs) derive ZERO half-states leads; the undetermined bucket stayed at 35. No changeset — workflow plus a root-private script entry releases nothing, matching the dispatch-gates wiring precedent that shipped the identical surface; `skip-changeset` applied and read back.",
 "tests": "RED BASELINE FIRST: `node scripts/pm/check-half-states.mjs --self-test` on the branch base -> exit 0, '✓ check-half-states self-test: 79 cases pass.' (green BEFORE wiring, so a red after wiring is the wiring). REVERSE VERIFICATION, direction predicted before running (new gate red, pre-existing coverage green — because the card's whole claim is that the existing coverage is bounded to H7): fix committed first, then ablated `h3QueueAndDispatched` (&& to ||) -> `pnpm check:pm-half-states` exit 1, '✗ check-half-states self-test: 1 of 79 case(s) failed.'; `pnpm check:partof-closing-keyword` exit 0, '✓ check-partof-closing-keyword self-test: 28 cases pass.' Both halves as predicted — the new gate has teeth AND the incidental H7 coverage is blind to H3, so the gap is real. Restored via `git checkout <branch> -- <path>` from the commit; 79/79 green again, tree clean. No ablation-on-dist involved (pure node script, nothing resolved from dist). GATE UNION (prompt list re-derived against the actual diff via `node scripts/pm/dispatch-gates.mjs package.json .github/workflows/lint.yml`; the re-derivation added NOTHING the prompt had missed — derived set {node-version, required-contexts, shard-attestation, type-check-coverage, type-check-debt, workflow-status-functions, check-shard-attestation.mjs} is a strict subset of the dispatch list). All exit 0: check:node-version, check:required-contexts, check:shard-attestation, check:workflow-status-functions, check:nul-bytes (7688 files scanned), check:filter-alias-parity, check:type-source-resolution, check:changeset-gate-self-tests, check:pm-dispatch-gates (101 cases), check:pm-half-states (79 cases), node scripts/check-changeset-no-major.mjs, node scripts/check-empty-changeset.mjs, node scripts/check-shard-attestation.mjs, check:type-check-coverage (64/77 packages). check:type-check-debt DID demand the built closure as the prompt warned — built it exactly as lint.yml does (`turbo run build --filter='./packages/*' --filter='./packages/*/*'`, 70/70 tasks, 6m46s, under flock + heap cap), then exit 0: '33 ledger entr(ies) re-measured in 250.4s, 1969 raw tsc error(s) total, none above its recorded number.' Remote CI in_progress at report time per the report-at-draft-PR-time rule; the PR-scoped 'Part-of PR must not also close its card' guard already returned success.",
 "open_questions": [],
 "out_of_scope_findings": []
}

Notes for the PM, outside the JSON:

• PM mechanism assumptions — all three verified, none falsified. (1) The self-test was green on main before wiring. (2) The identity-hint derivation that landed this hour works end-to-end: node scripts/pm/dispatch-gates.mjs scripts/pm/check-half-states.mjs now prints pnpm check:pm-half-states [lint.yml] matched via scripts/pm/check-half-states.mjs, so cards editing that file derive the new gate automatically. (3) The setup-node trap did not bite — a step was added to an existing job and no setup step was touched.
• Labels read back after the bots settled: ci/cd, dependencies, size/s, skip-changeset. My skip-changeset write survived; Auto Label added the first two on top of it rather than wiping it.
• packages/spec gate-to-generator ledger is not implicated — it reconciles that package's own package.json (pkgRoot = packages/spec), not the root's, so a new root check: entry is out of its scope. Consistent with the dispatch-gates wiring precedent, which added a root entry and touched no ledger.

---

Generated by Claude Code

---

Generated by Claude Code

[hotlong]

Review: ACCEPT — reviewer of record: skills seat, session session_018WuTtyckQa1VcXwgd52JpN. PR #8544 verified behaviorally at its head (79/79 through the new entry; the workflow step unconditional beside its family; +36 across exactly the two claimed files).

• The measured decision is the review's centerpiece: direct entry, no thin gate file — opposite of the dispatch-gates answer and correct, because it was measured under the landed masking (1 module-body hint, the repo slug, unpairable with any repo path) instead of copied. Eight probe paths derive zero false leads; census 93→94; bucket preserved at 35.
• The reverse verification proves the card's own claim: ablating H3 reddens the new gate while the blocking guard's 28 cases stay green — the incidental H7 coverage really was blind to the other six predicates, so the wired gate closes a real gap, not a formality.
• Bonus verified at review: a card editing check-half-states.mjs now derives BOTH its own gate (identity) and the blocking guard (via its import specifier as a module-body hint) — the dependency chain the finding worried about is now visible to dispatch derivation end-to-end.
• Discipline: green-before-wiring baseline; check:type-check-debt built the closure rather than arguing it away; re-derivation a strict subset of the dispatch list; skip-changeset per the identical-surface precedent.

Landing: normal queue path. Ready-flip + auto-merge once the gate jobs on head 3fbf4bf read success — flip check armed. This closes the unwired-self-test family's fourth and last known member.

---

Generated by Claude Code