[finding] 10 governed skill fences re-declare a published @object-ui type instead of importing it — the fence-repair half of objectui#7646, which the gate PR could not touch #8335

Filed by the domain:devx @ objectui execution seat from the objectui#7646 implementation run (R46). Measured on fedfa3e4a, not inherited — objectui#7646's own numbers were re-derived rather than copied.

objectui#7646 asked for the gate-design half and got it: scripts/check-skill-examples.mjs now carries a fourth assertion — a marked (os:check) fence that locally re-declares a name the built @object-ui/* surface publishes must either import that name or carry a row in a shrink-only ledger declaring it a deliberately simplified teaching copy. This card is the other half: the fence repairs themselves.

Why they are not in that PR

skills/** and .claude/skills/** are GOVERNED surfaces — agent drafts, human merges (scripts/check-governed-queue-guard.mjs, GOVERNED_SURFACES). The gate PR touches scripts/ only and every path in it comes back NOT GOVERNED. A PR that edited a fence would park for a human merge and strand the gate change with it, so today's offenders were landed as ledger rows instead. Each row is declared debt, never a green — the run prints the row count on every pass.

The population, re-derived on fedfa3e4a

Census over every ts / tsx / typescript fence in skills/, .claude/skills/ and content/docs/, matching each fence's local type NAME / interface NAME against the exported type and interface names of the workspace packages, excluding any name the same fence imports:

• 602 fences in the corpus, 13 carry os:check.
• 28 re-declare a name the workspace exports; 26 distinct names.
• Split by root: skills/ 8, .claude/skills/ 2, content/docs/ 18.
• Lit control: the census's own marked count (13) equals what check:skill-examples prints for itself on the same tree (Marked: 13 ts fence(s), Semantic phase: 13 of 13 ts fence(s) judged). The instrument was not blind.

This card owns the 10 under the two governed roots. MARKED rows are the ones the new assertion gates today and the ones carrying a ledger row.

What a repair is, per fence — ⛔ not a sweep

objectui#7646's triage ruled option 3 (blanket conversion to imports) out: a guide's fence is often a deliberately simplified view of a large published type, and a mechanical pass would make several guides less readable. So each row is one of three judgements, made by the guide's owner:

1. Import it. import type { AuthUser } from '@object-ui/auth'; and delete the local copy. Drift becomes a compile error at the point it happens. Best when the fence's copy is already the whole type.
2. Derive from it. Import the published name under an alias — import type { QueryResult as PublishedQueryResult } from '@object-ui/types'; — then define the guide's short shape from it with Pick or Omit. Keeps the fence short AND anchored, and the new assertion accepts it: a name the fence imports is not a finding under either spelling, the alias or the original.
3. Keep the teaching copy and say so. Leave the fence, and leave its ledger row in KNOWN_SHADOWED_PUBLISHED_TYPES with the reason. ⛔ The row is uncovered debt, not a pass — the gate's summary says so on every run.

Retiring a row and editing its fence must land in the same commit: a row whose fence now imports the name fails as STALE, exactly like this gate's two older shrink-only lists.

The evidence this is worth paying for

Reproduced on fedfa3e4a with trap/restore (mutation confirmed on disk by anchored counts; restore proven by blob hash equal to the HEAD blob and an empty git diff HEAD), on fence 6:

• Arm B — an internally CONSISTENT lie: name: number where the published required member is string, all five ADR-0049 retirement tombstones restored as writable, plus an invented frobnicate that exists on no type at all. check:skill-examples exits 0, printing Semantic phase: 13 of 13 ts fence(s) judged, 0 failed. The fence is INSIDE the judged count — the gate read it and passed it.
• The recorded miss, kept because it is informative: an earlier arm that left inputType?: never in place and added inputType?: string went RED with TS2300 (twice) and TS2717. That is the gate catching an internally INVALID snippet, which is exactly its job, and says nothing about drift. ⛔ Anyone re-measuring must plant an internally consistent lie or they will conclude the gate works.

⛔ Not in this card's scope

• The 18 content/docs/ fences. check:doc-snippets is opt-OUT, so those 18 ARE compiled and equally uncomparable — the blind spot spans both gates. Widening the assertion into scripts/check-doc-snippet-types.mjs is a separate gate-design card against a much larger population, and it is ⛔ not started here. It also lands inside that script's licensed strictness region, which objectui#7646's dispatch put off limits.
• The gate itself. Landed in the objectui#7646 PR; this card changes no script.

Merge posture

⚠️ Every path this card touches is GOVERNED. Its PR is agent drafts, human merges: it parks in draft for a maintainer, ⛔ is not flipped ready by a seat, and ⛔ does not self-queue.

Back-links: objectui#7646 (the finding and the gate), objectui#7636 (the instance that paid for it first).

Generated by Claude Code