Two-legged identifier-liveness gate over the published skills catalog (#13678) (#13892)

* feat(devx): two-legged identifier-liveness gate over the published skills catalog

Leg 1 (phantom row): a skills/** table row citing an identifier that greps to
zero in packages/** is a red. Leg 2 (missing row): a section REGISTERED as
exhaustive over a schema symbol that does not document every member is a red.
One corpus walk, two legs.

Leg 2 is table-driven rather than inferred, and that is a measurement result:
auto-binding a doc table to any enum it is a subset of scores 1 true positive
in 3 on this corpus. The heuristic ships as --suggest, which cannot fail.

Zero skills/** edits: the five surviving Leg 1 sites and the one measured Leg 2
gap are ledgered, not fixed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EnE7G31tqbxN1rqpQmzurT

* chore(devx): register the liveness gate with the farm's convention gates

- check-ratchet-remedy-authority CONTROL row (the sweep classified it `marked`
  before the row was written; the row is recorded from that verdict).
- bare-root-worklist: REFUSE-WIDE for IMPL_ROOTS `packages` — the word index
  reads 5405 of 5640 tracked files (96%), so the population IS the root and a
  true declaration would name this gate on every card touching any package.
- The watch-hint self-test asserts the spelling without writing a second
  declaration site; spelling it as a regex literal made the real one
  unjudgeable ("2 declaration sites").

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EnE7G31tqbxN1rqpQmzurT

---------

Co-authored-by: Claude <noreply@anthropic.com>

Author: huangyiirene

.github/workflows/lint.yml

@@ -1862,6 +1862,38 @@ jobs:
       - name: Teaching-corpus lexical anti-drift ratchet
         run: pnpm check:corpus-claim-drift
 
+      # #13678: the third gate over this corpus, and the only one asking an
+      # EXISTENCE question rather than a lexical one. The two steps above ask
+      # whether a forbidden word appears, and whether a spelling stands next to
+      # prose naming a semantic it does not have. This one asks, in both
+      # directions, whether the identifiers the catalog names and the schema
+      # members it enumerates are the same set:
+      #
+      #   Leg 1 (phantom row)  a table row citing an identifier that greps to
+      #                        zero in `packages/**` is a red. Measured: 3 of 6
+      #                        falsehoods in one skill were this class, every one
+      #                        catchable by a single grep of the cited name.
+      #   Leg 2 (missing row)  a section REGISTERED as exhaustive over a schema
+      #                        symbol that does not document every member is a
+      #                        red. Measured on a later flight: 15 of 20.
+      #
+      # Leg 2 is table-driven rather than inferred, and that is a measurement
+      # result, not a preference — auto-binding a doc table to any enum it is a
+      # subset of scored 1 true positive in 3 on this corpus, because deliberate
+      # subsets ("Three action types dispatch headlessly") and semantically mixed
+      # enums are indistinguishable from real gaps without a human. The heuristic
+      # ships as `--suggest`, which cannot fail anything.
+      #
+      # It lives in this job for the reason every other docs guard does
+      # (`check:doc-anchors`, `check:role-word`, `check:corpus-claim-drift`
+      # above): this job carries the whole `check:*` gate family and has no paths
+      # filter. That matters more here than for its neighbours — the gate's two
+      # inputs are `skills/**` and `packages/**`, so a paths filter on either
+      # would let the OTHER side drift a member in or an identifier out without
+      # ever running the check that notices.
+      - name: Published-skills identifier liveness gate (two legs)
+        run: pnpm check:skill-identifier-liveness
+
       # #11671: `os i18n extract --fill=default` fills GAPS only, so revising a
       # source string rewrites `en` and strands the previous source text in every
       # other locale. The bundle stays in sync BY KEY, so `check:i18n` reports OK

package.json

@@ -55,6 +55,7 @@
     "check:template-version-sync": "node scripts/sync-template-versions.mjs --self-test",
     "check:role-word": "node scripts/check-role-word.mjs --self-test && node scripts/check-role-word.mjs",
     "check:corpus-claim-drift": "node scripts/check-corpus-claim-drift.mjs --self-test && node scripts/check-corpus-claim-drift.mjs",
+    "check:skill-identifier-liveness": "node scripts/check-skill-identifier-liveness.mjs --self-test && node scripts/check-skill-identifier-liveness.mjs",
     "check:quick-reference-counts": "node scripts/check-quick-reference-counts.mjs --self-test && node scripts/check-quick-reference-counts.mjs",
     "check:overlay-whitelist-table": "node scripts/check-overlay-whitelist-table.mjs --self-test && node scripts/check-overlay-whitelist-table.mjs",
     "check:runtime-services-index": "node scripts/check-runtime-services-index.mjs --self-test && node scripts/check-runtime-services-index.mjs",

scripts/check-ratchet-remedy-authority.mjs

@@ -756,6 +756,34 @@ const CONTROL = {
     why: 'PR for #13582. Its per-file, per-rule baseline is shrink-only; the author\'s remedy is to reword the sentence, and the baseline path that admits a genuinely true construction beside a pinned spelling is offered only as a maintainer\'s.',
   },
 
+  // The eleventh, and the third gate over the published catalog (#13678). It
+  // carries TWO offers on ONE ledger file and both are marked, which is why the
+  // row is worth reading rather than counting:
+  //
+  //   LEG 1 offers an EXEMPTION — "this identifier is live somewhere this repo
+  //   cannot index". That half is shrink-only BY CONSTRUCTION rather than by
+  //   policy: `--update` prunes entries the scan no longer reaches and there is
+  //   no code path that adds one, because the `kind`/`note` a row must carry are
+  //   human judgements nothing in the tree encodes. A hand edit is therefore the
+  //   only way in, and it is the maintainer's.
+  //
+  //   LEG 2 offers a GAP entry — a measured count of schema members a section
+  //   registered as exhaustive does not document. That half IS regenerated from
+  //   the tree, so its flag moves whichever way the tree moved and only policy
+  //   tells a ratchet-down from a weakening apart. Same treatment as the two
+  //   `--update` gates above.
+  //
+  // Both messages offer the author's own remedy first and unmarked (correct the
+  // row; document the member), which is the shape this convention wants.
+  //
+  // ⚠️ Recorded from the sweep's own verdict — this gate reported it as `marked`
+  // before the row was written, and the row was written because the sweep
+  // demanded it. Author intent and detector agreeing is the point of the corpus.
+  'check-skill-identifier-liveness.mjs': {
+    expect: 'marked',
+    why: 'PR for #13678. A two-legged liveness gate over skills/**: a row citing an identifier that greps to zero, and a section registered exhaustive that omits a live schema member. One ledger, two halves — the Leg 1 exemptions are shrink-only by construction (--update only prunes them), the Leg 2 gaps are a regenerated shrink-only ratchet — and both weakening paths are the maintainer\'s, offered after the author\'s own.',
+  },
+
   // ── Declaration registries and near-misses: recording the fact IS the fix ──
   'check-agent-model-declared.mjs': {
     expect: 'excluded',