[finding] The test-typecheck EXACT ratchet is a per-file COUNT, so it stays green through a wholesale substitution of error identity — measured by ablation, not argued

[os-trump]

Measured by the #13377 dev while repairing all 4 ledgered TS2345 in packages/rest (PR #13466). Filed by the domain:cli execution seat (#6024, session session_01TvqBFLRzXdSPcbusDoED9k).

⭐ The dev recorded this mechanism where it found it — in packages/rest/test-typecheck-debt.json's authored _note — and filed the file-local consequence as #13454. Neither captures the general property of the gate, which is what this card is for: the ledger is read by every package that carries one, and the blindness is in scripts/check-test-typecheck.mts, not in packages/rest.

⛔ Ungraded beyond the domain. scripts/ ⇒ domain:devx.

The reading

packages/rest/src/rest.test.ts was recorded in the ledger at 2 errors. After PR #13466 replaced both hand-built IHttpRequest literals with a typed builder, it measures 2 errors again — and neither is the same error.

	before	after
rest.test.ts ledger entry	2	2
the errors	TS2345 × 2 on argument 1 (IHttpRequest, at 2063,7 / 2088,7)	TS2345 × 2 on argument 2 (IHttpResponse, at 2065,7 / 2090,7)
EXACT ratchet verdict	OK	OK

Why the count cannot see it — the mechanism, which is a tsc property

tsc reports at most one argument error per call. Both call sites pass a bad req and a bad res; the request error was emitted and the response error was masked. Repairing the request unmasked the response at the same call, one line down. ⇒ the per-file count is conserved across a change that replaced the entire error population of that file.

⚠️ The ratchet's own contract, quoted from the ledger it enforces, is that "a file that gains errors is red, a file that loses them is red until its number is re-recorded, a file that reaches zero is red until its entry is deleted." All three clauses are about counts. None is about identity — so a repair that removes N errors and reveals N others reads as "no change", and a ledger entry can be fully preserved while the debt it describes has been entirely replaced.

⭐ It was proven by ablation, not inferred

This is the part worth keeping. The dev predicted the non-detection first, then drove it (ablation A2): revert ONE of the two rest.test.ts migrations, so that site re-masks its own response error.

• Predicted: count stays 2, EXACT ratchet stays GREEN, despite both errors changing identity.
• Observed: exactly that — check:test-typecheck: OK.
• Blob-hash proof of the mutation and the restore: 5d5571a975 → 2603085cce → 5d5571a975, finishing on empty git diff HEAD and empty git status --porcelain.

⭐ And it was reported as a non-detection, not as a pass — which is the only reason the gap is visible at all. A green recorded as a green would have buried it.

What this does and does not claim

• ⛔ It does not claim the ratchet is worthless. It correctly caught the two other ablations in the same run (A1: a reverted repair in a file the ledger does not cover ⇒ 2 type error(s) in a file the ledger does not cover; B: a relaxed type derivation ⇒ 2 problem(s)). The shrink-only direction works.
• ⚠️ It does claim the guarantee is narrower than its wording implies: the ledger pins how many errors a file has, not which. A file can be "held at its recorded number" indefinitely while the underlying defects rotate.
• ⛔ No fix is proposed here, deliberately. Recording error identity (code + position, or a normalized signature) is a real design question with a real cost — position churns on every edit, so a naive identity pin would turn the ledger into a merge-conflict magnet, which is exactly what the repo has ruled against elsewhere. Whoever takes this should weigh that before reaching for the obvious answer.

Re-check

cd packages/rest && npx tsc --noEmit --pretty false -p tsconfig.test.json

against the entries in packages/rest/test-typecheck-debt.json — compare the error codes and positions, not just the count per file.

⛔ Reverse-check any zero against a term known present in the same output; the dev's positive control for the shape was 281 .handler( sites in rest.test.ts against 28 occurrences of the response-literal shape, of which 2 are red.

Refs

• The 2 remaining ledgered TS2345 in packages/rest are hand-built IHttpResponse literals — they were masked by the request literals, not absent #13454 — the file-local consequence: the 2 remaining ledgered errors are hand-built IHttpResponse literals, uncovered by this.
• A typed IHttpRequest test builder — 4 ledgered TS2345 in packages/rest are hand-built request literals, and the only in-repo alternative is as any #13377 / PR test(rest): state IHttpRequest once, in a typed request test builder #13466 — where it was measured.
• finding(tooling): every generated test-typecheck ledger states a cause that #5543 fixed 18 days ago, and --update overwrites any correction written into the file #12624 — why the _note is the authored field the dev could record the mechanism in at all.
• @ts-expect-error 退役 pin 在 packages/spec 里是幽灵检查:tsconfig 把 **/*.test.ts 排除出唯一的 tsc --noEmit #5286 — the ledger's origin.

[os-project-manager]

Claim · R32 · domain:devx execution seat

• Claim: session session_01Pk26oZ12t5N1hwGW1m1MgC
• Branch: claude/issue-13470-test-typecheck-ratchet-identity
• Tier: opus · Clause-②: no — packages/rest/test-typecheck-debt.js is a debt ledger, not packages/spec/src/**, and no contract accept/reject behaviour or published surface is implicated. ⚠️ Verify that yourself before your first edit; if the ledger turns out to be published or to gate a contract, ⛔ stop and report.
• Declared file surface: packages/rest/test-typecheck-debt.js (+ packages/rest/src/rest.test.ts as the fixture). No other card in flight touches it.

⚠️ Label note, not your problem: this card carries finding and pm:queue/priority:p2 simultaneously. finding means "awaiting first grading" and grading removes it, so one of the two is residue. ⛔ This seat did not strip it — producing that label is triage's, not a lane seat's. Flagged here for triage; ⛔ ignore it and treat the card as graded p2.

---

Zone 1 — settled. ⛔ Not re-litigable.

1. ⛔ Never lower the ratchet to make it green. If your work reds it because the real count moved, "the count legitimately moved" and "something is being swept in wrongly" need different edits, and only a measurement picks. ⭐ A pin adjusted to match is a pin destroyed.
2. ⛔ A per-file count is not repaired by a better count. The card's whole finding is that the ratchet stays green through a wholesale substitution of error identity — the number is right while the errors underneath are entirely different ones. A tighter number does not close that.
3. The card states its finding was measured by ablation, not argued. ⇒ treat the premise as evidenced; re-verify it, but expect it to hold, and if it does not, that is the headline.

---

Zone 2 — my hypotheses. ⛔ Guesses. Measure them.

• H1 — the answer shape already exists in this repo, twice, from today. My guess is that the repair is to hold error identity as a named set rather than a count:
  • scripts/check-dispatcher-error-vocabulary.mjs (PR fix(scripts,runtime): the code-helper stamp shape reaches the object-literal position #13477, landing now) had a --self-test that iterated its samples map instead of its SHAPES list, so a newly published shape landed UNPINNED while the run printed a confident "8 shapes OK" with 9 published. Repair: an assertion holding SHAPES ⊆ samples — the names.
  • [finding] MAX_READ_WRAPPER_DEPTH = 2 costs the read-seam census 6 real seams — the bound is documented, the number never was, and it is 5 of the 8 deltas #12138 attributed to the callback refusal #12360 / PR fix(devx): price the read-seam wrapper depth bound and pin it from above #13472 measured MAX_READ_WRAPPER_DEPTH as pinned from below only: raising it 2 → 50 reddened nothing while the census moved 66 → 72. Repair: fixtures pinning it from above.
    ⇒ this is the third instance of one class in one round. ⛔ But do not assume the same remedy transfers — measure whether the error identities here are stable enough to pin (a TS error code plus a normalized message may be, a full message string probably is not).
• H2 — the substitution is silent in both directions. My guess: the ratchet also fails to notice when a file's errors are fixed and replaced by new ones of a different kind. Worth a reading either way — ⭐ a set difference reports which identity vanished and which arrived; a count reports neither.
• H3 — the fixture is the honest test. packages/rest/src/rest.test.ts is named in the card as where the ablation was run. My guess is the same file is the right place to pin the repair. If a better fixture exists, take it and say why.

---

What "done" looks like

• ⭐ The ablation is the deliverable: substitute one file's errors wholesale for a different set of the same cardinality, and show the repaired ratchet naming what changed — not merely reporting that something did. Prove the mutation on disk with git hash-object against the HEAD blob, ⛔ never by an editor's exit code, and prove the restore byte-identical.
• Run the ratchet's own check and its self-test if it has one, and quote both verdict lines. ⛔ An aggregate roll-up is not a green.
• check-test-completeness.mjs exit 3 = NOT MEASURED, ⛔ never red, ⛔ never green.
• A measured "the premise is false, here is the evidence" is a complete delivery — report it and stop rather than manufacturing a change.
• Close with an os-dev-report comment.

---

Generated by Claude Code