Gate candidate: named-identifier liveness check over skills/** — a skill row citing a schema/field that greps to zero repo-wide is a red, not a sentence

[zhuangjianguo]

Filed from flight ①'s calibration (program #13658, PR #13674): 3 of the 6 measured falsehoods in objectstack-formula were PHANTOM ROWS — surface tables citing identifiers that no longer exist or never did (etl.schedule / sync.schedule retired #6414/#4738 · system/notification's 5 template fields removed under ADR-0049 #4610 · titleTemplate/bodyTemplate with zero hits across packages/**/*.ts). Each was detectable by one repo-wide grep of the cited identifier — no semantics needed.

The gate's shape (dispatch decides details; this is the measured direction)

A scripts/check-skills-* gate that extracts identifier-shaped citations from skills/** table rows (backtick-quoted schema keys, field names, metadata kinds) and greps each against the implementation tree: zero hits = red naming the row and the identifier. Known hard parts to design for, from flight ①'s data:

• allowlist for deliberately-abstract placeholders (<task>, record.x) and for prose examples that are not citations;
• a positive control in the self-test (a known-live identifier must be found; a known-phantom seeded fixture must go red) — the instrument-calibration discipline;
• it complements, never replaces, the behavioral sweep: it catches existence, not semantics (defaultValue mis-typed as an Expression slot greps fine and is still false).

Why it clears the value gate

Mechanical construct over prose vigilance (#12911); would have caught 3/6 of flight ①'s falsehoods for pennies; the sweep program (#13658) fixes today's corpus but nothing stops tomorrow's retirement PRs from re-creating phantoms — ADR-0049 removals are routine and skills tables are exactly where their corpses accumulate. Runs in the existing skills gate family beside the token ratchet.

Serial note: independent file surface (new script + CI wiring) — no conflict with the sweep flights. Refs: #13658 (program) · PR #13674 (the three phantoms, with evidence) · PR #13577 (the class's first proof).

Generated by Claude Code

[huangyiirene]

Scope note from the lane seat (skills-lane self-triage; no re-grade, card stays pm:queue): flight ⑧ of the factual sweep (#13815 / PR #13833) measured the INVERSE of this card's phantom-row direction as the dominant falsehood class — 15 of its 20 false landing sites were enumerations that stopped growing when the schema did (a live schema member with NO doc row: _tabs, pages, flows, setSupportedLocales, 9-of-14 CLI kinds…). When this gate is built, it should carry BOTH directions over the same corpus: (1) this card's original leg — a skill row citing an identifier that greps to zero repo-wide is a red; (2) the inverse leg — a doc enumeration presented as exhaustive whose row set is a strict subset of the schema's member set is a red (the presented-as-exhaustive predicate needs care: e.g. lists are not falsifiable by omission, per flight ⑥'s recorded reading). One gate, two legs, one corpus walk. Evidence trail: #13815's report; flight ⑤ found the same shape (10-member enum, 9 rows; 20-member ChartType, 19 rows); flight ⑦ (driver catalog missing two live ids). Serial note stands: this card still waits on PR #13584's .github/workflows/lint.yml face before dispatch.

---

Generated by Claude Code