fix(pm): name the ratchet invocation CI runs in dispatch-gates' test-file entry

[hotlong]

Fixes#8545
Fixes#8604

No gate is weakened and no invocation is removed from CI. What changes is what the DERIVATION prints.

The mismatch

CHANGE_KIND_GATES's test-file entry emitted check:type-check-coverage under a rationale describing the TEST_DEBT ratchet. The ratchet is the --re-measure half, and only check:type-check-debt passes it — both resolve to one script, scripts/check-type-check-coverage.mjs. Measured on the merged ref: lint.yml runs pnpm check:type-check-coverage at line 1021 and pnpm check:type-check-debt at line 1449, the latter immediately after a pnpm exec turbo run build step that builds the ledgered packages' closure.

What the entry emits now

Both invocations, each stating the question it answers:

• check:type-check-coverage — the structural half. A package whose test files sit outside every tsc program accounting for it must carry a TEST_DEBT entry, so a new test file no tsconfig reaches moves it. It re-measures no count.
• check:type-check-debt — the ratchet half, and the invocation CI runs for it. Carries the built-closure prerequisite inline, with the exact build command lint.yml runs before that step, and says the cold-worktree refusal means NOT MEASURED, never "not applicable to me".

A new How these entries stay honest bullet writes down the general rule: every name in that table is an INVOCATION, not a script.

The class sweep the card asked for

Swept over the merged ref: the workflows discover 96 check families resolving to 73 distinct script files; 8 of those files are reached by more than one family. Seven of the eight are a different shape — a root check: script beside a direct node scripts/check-x.mjs step in a second workflow (the three changeset gates, cross-package-test-inputs, objectui-pin-fresh, partof-closing-keyword, shard-attestation). The derivation already discovers each of those direct invocations as its own family and prints it with its own runnable invocation, so nothing is misdirected there. check-type-check-coverage.mjs is the only file whose two CI invocations are two ROOT SCRIPTS differing by a flag, and the only one this table names. One-off today; the rule is what generalises, not the fix.

Pins

The self-test grows from 103 to 106 cases. All three convention gates are pinned on the rendered delimiters (- pnpm x —), never a bare substring, and the STALE branch is pinned per NAME (⚠ x: STALE) so the pair sharing one script is reported apart. One further case holds the ratchet line's prerequisite text (the built-closure condition plus a runnable turbo run build), because an unexplained throw is what reads as "not applicable".

Reverse verification, direction predicted before running, from the committed state:

• rename both gates to -v2: exactly the 4 anchored cases go red, both COUNT pins stay green (that blindness is why the anchors exist), and the live run prints both names as STALE — the STALE branch works for both printed names.
• delete the ratchet entry (the pre-fix shape): 5 cases red, including both count pins.

Restored from the commit afterwards, byte-identity confirmed with git hash-object against HEAD:scripts/pm/dispatch-gates.mjs.

Rider, #8604 — comment-only, no executable change

scripts/pm/check-dispatch-gates.mjs's header justified its unquoted repo paths with "watch-hint extraction ... does not skip comments". extractWatchHints has opened with maskComments since #8478, and that function's docblock names this file as the specimen it retires.

Measured with extractWatchHints, every repo path in the header rewritten into backticks: 1 hint under today's extractor (the module-body constant, the same single hint the file ships with) against 10 under the pre-masking extractor, reaching spec, objectql, plugins, drivers, .claude/agents and .changeset exactly as the incident describes. Comment masking alone accounts for the difference; masking self-test bodies instead changes nothing here, because this file has none. So the premise holds: the convention is no longer load-bearing. The incident is kept as history, the ⚠️ instruction line is dropped, and the rule that survives is stated — a module-body literal is what masking cannot reach.

The separate-gate-file argument in the same header gets today's numbers too: the tool's own source yields 4 hints, not the 49 the header records, and three of them (packages/plugins, packages/drivers, packages/services) are directory bases the package resolver really probes and wide enough that a directly-wired gate would still fabricate leads — so that decision stands, on narrower grounds than the text claimed.

One measured detail worth keeping: the ratchet line's remedy command spells its --filter values unquoted (and says to quote them for the shell) because a why string is a MODULE-BODY literal that masking cannot reach — the shell-quoted spelling adds both glob values to this file's own hint set, inert today only because hintCovers rejects a hint collapsing to a bare top-level directory.

Verification

• pnpm check:pm-dispatch-gates — 106/106 pass
• pnpm check:nul-bytes — OK, 7757 files scanned; plus a targeted control-byte self-scan of both changed files
• pnpm exec eslint on both changed files — clean
• re-derived with node scripts/pm/dispatch-gates.mjs against the actual changed paths: check:pm-dispatch-gates only, nothing beyond the dispatched list

Dev tooling only (scripts/pm/), releases nothing — skip-changeset.

Related: #8442 / #8533 (where the misdirection cost a CI lap), #8478 (the masking), #6376 (why --re-measure demands a built closure), #5278 (the shrink-only rule). #8538 is a different finding about the same script family and remains open; it is not addressed here.

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 14, 2026 1:57am