tooling(scripts): assertion floors for the batch 6a self-tests — sink repair plus one hoisted battery (#13799 batch 6a)

[claude]

Part of #13799 (batch 6a — the class-1 + class-2 deferrals from PR #15003's table whose sink is N inline writes or a concise-arrow helper, and whose named section banners number fewer than two).

Dispatched by PM claim 5536968867; dev claim 5536986594. Concurrent with batch 5 (PR #15217, since landed) and batch 6b — file sets disjoint, verified by set intersection.

failures.length === 0 / failed === 0 was these self-tests' only success condition, so "every case held" and "the cases never ran" printed the same line. Eight of the nine candidate files now carry PR #13487's roster-and-floor shape; the ninth is deferred with a measured reason.

What each file got

Two recipes, both landed and both applied verbatim:

• the sink repair (PR test(scripts): give six more self-tests a battery roster and floor (#13799 batch 3) #15156, batch 3) — where the sink is a per-case concise arrow, it gains a block body that calls registerCase() before the unchanged push;
• the thunk (PR test(scripts): floor partition-test-shards self-test via the thunk route (#13799 batch 4) #15198, batch 4) — where the inline sink writes are failure-only, check(() => { ... }) registers the case and then runs the existing site verbatim. This is the route the card names for sites that would otherwise need rewriting, and it is the only honest one here: routing failure-only writes through a helper would floor zero cases on a green run, which is no floor at all. Registration happens whether or not the site fires.

⛔ No assertion condition is inverted or rewritten anywhere in this diff. ⛔ No comment was promoted to a section head; the two files carrying a single named banner were not split on it.

file	sink at the pin	repair	battery (roster size pinned at 1)	floor
scripts/check-console-intercept-disarm.mjs	failures counter, 2 inline sites	thunk × 2	check-console-intercept-disarm self-test	11
scripts/check-declaration-mirrors.mjs	concise-arrow ok	block body + registerCase()	check-declaration-mirrors self-test	23
scripts/check-docs-single-h1.mjs	concise-arrow t	block body + registerCase()	check-docs-single-h1 self-test	20
scripts/check-error-code-casing.mjs	failed counter, 3 inline sites	thunk × 3	check-error-code-casing self-test	51
scripts/check-examples-live-imports.mjs	failed counter, 1 inline site	thunk × 1	check-examples-live-imports self-test	36
scripts/check-meta-type-normalized.mjs	problems array, 4 inline sites	thunk × 4	check-meta-type-normalized self-test	4
scripts/check-optional-error-sink-contract.mjs	failures counter, 7 inline sites	thunk × 5 (see below)	check-optional-error-sink-contract self-test	66
scripts/check-org-identifier.mjs	failed counter, 2 inline sites	thunk × 2	check-org-identifier self-test	31

Every floor was measured, never transcribed: the roster was pinned to an unreachable value, the self-test run, and its own breach line (registered N case(s), below its pinned floor of ...) read for N. Four of the eight print a case count of their own, and all four agree with the measurement — check-org-identifier 31 = its cases.length + 1, check-error-code-casing 51 = its printed 46 + 5, check-examples-live-imports 36 = its cases.length, check-console-intercept-disarm 11 = its printed 10 + the real-tree floor, and check-optional-error-sink-contract 66 = its printed 19 + 9×4 spellings + the reject side + 9 derivation rows + the run reference pin.

One deliberate non-registration

check-optional-error-sink-contract's two assertions over baseline.entries are left unregistered on purpose, and say so in place. They run one-per-row of a ⛔ shrink-only ledger, and a floor moving with that list would red every legitimate removal — training the next author to edit the floor, which is the one habit these floors exist to prevent (#13797's ruling, carried forward by PR #15217's check-whole-set-label-write). LOG_CHANNELS is not that: it is the contract's own vocabulary rather than a list meant to shrink, so its two loops do register.

Deferred, with a measurement

file

measured reason

scripts/check-agent-model-declared.mjs

Its failed sink has 7 inline sites, and 3 of them cannot be wrapped verbatim: all three sit inside the for (const c of cases) loop and terminate in continue, which cannot cross a thunk's function boundary, and one of the three is a catch clause rather than an if statement at all. There is no per-case helper to give a block body to. The 4 discovery-block sites outside the loop are thunk-able, but flooring only those would pin 4 of 22 cases (18 table rows + 4 discovery) while flipping the file's census class from NONE to ROSTER — the census would then report a hole as closed that is mostly still open, which is worse than an honest NONE (PR #15217's ruling, restated). Deferred rather than improvised, per the card's own instruction.

The residue for this file is a real class decision, not a one-file judgement: what to do with a table-driven self-test whose per-case decision is made by continue in a loop body. It belongs with check-startup-registry-verdict and js-comment-mask, which PR #15217 deferred for the neighbouring reason.

Evidence

• Case count before == after, measured. Roster pinned to an unreachable value on each file, breach line read for the registered count; restored and re-verified. --self-test stdout and stderr byte-identical to the base tree's on all eight, exit 0 on both sides; normal mode also byte-identical on the three that have one.

• Line-level identity of the sink rewrite. Over the whole diff, 68 removed lines against 937 added: after whitespace normalisation exactly 2 removed lines do not reappear, and both are the concise arrows that gained block bodies (const ok = (label, cond) => cases.push(...) and const t = (name, ok, detail) => cases.push(...)). Their expressions reappear verbatim as their own lines — the residue is the arrow head becoming a block head, nothing else.

• Ablations, each from the committed implementation, each proving the mutation on disk by an anchored count and a git hash-object that moved off the HEAD blob (an empty hash read as FAILURE, not "nothing to compare"), each proving the restore by hash equality against the HEAD blob and an empty git diff HEAD — never by an exit code, never by a bare git checkout -- path (that reads the index the mutation also wrote). Anchors had to match exactly once or the leg aborted; a deletion leg additionally proves the byte length moved by exactly the anchor's length. No dist is on the resolution path of any of the eight, so no rebuild leg applies.

  • battery opener renamed, all eight: exit 0 → 1 on every one, each naming itself, e.g. self-test battery "check-optional-error-sink-contract self-test" DID NOT RUN — 0 cases registered, 66 pinned.
  • the single roster row deleted (check-org-identifier): exit 0 → 1 naming both halves — SELF_TEST_BATTERIES declares 0 batteries, below the pinned 1 and registered 31 case(s) but is not declared in SELF_TEST_BATTERIES — so the size pin is load-bearing even at size 1.
  • the introduced registerCase() removed from the sink helper (both shapes: check-declaration-mirrors, check-docs-single-h1, check-error-code-casing): exit 0 → 1 with every original case still passing — zero original-case failures in the output — and the floor reporting DID NOT RUN.

• Census (node scripts/measure-self-test-floor.mjs --json), as set assertions rather than read by eye: ROSTER 113 → 121, COUNT 2 → 2, NONE 54 → 46. Unfloored NONE drops by exactly 8; the set of files whose class changed equals the worklist exactly (symmetric difference empty); every transition is NONE → ROSTER; git diff --name-only against the pin equals the worklist exactly.

• Watch hints: extractWatchHints run over each of the eight on both trees, both with a real watchHintTree bundle built from that tree's git ls-files and without one — 8/8 byte-identical. (A first harness passed tree as a path string and crashed identically on both sides; that "identical" was two identical crashes, not a measurement, and was discarded rather than reported.) Battery names contain a space, so they cannot pass the hint admission test at all.

• Gates. Derived with node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (no paths passed — it reads the change set from the merge base) = 31 commands; plus each edited script's own --self-test, plus the always-runs tail (check-self-test-wired both modes, check-self-test-workflow-commands.mjs both modes, check:declared-population-live, check:ratchet-remedy-authority, check:watch-hint-literal, check:nul-bytes). 44 run, 44 exit 0, every exit code captured by redirect before any pipe — including check:pm-dispatch-gates in the foreground (✓ dispatch-gates self-test: 1353 cases pass.). Nothing read NOT MEASURED. One invocation of mine was wrong rather than red: pnpm check:self-test-wired returned 254 with ERR_PNPM_RECURSIVE_EXEC_FIRST_FAIL Command "check:self-test-wired" not found — there is no such package script; the real gate is the node scripts/check-self-test-wired.mjs form the derivation itself printed, and it ran green in both modes. Recorded rather than counted as a failure.

• pnpm lint whole repo (eslint . --no-inline-config) exit 0 in 99 s — no narrowing claimed.

• Control-byte scan (grep -naP over the control range) across all eight changed files and every body posted to GitHub: no hits.

• All of the above ran on 3263b6955, this branch's head and its final commit, with git status --porcelain empty.

Notes

• skip-changeset: nothing is published from any package.
• The derivation printed a STALE TREE notice — origin/main advanced 4 commits (PR tooling(scripts): assertion floors for the class-2 self-tests, one hoisted battery each (#13799 batch 5) #15217 landing among them) and 11 files it derives from changed. The branch is deliberately not merged forward: none of those commits touches any of these eight (set intersection empty, merge-base still the dispatch pin 1bc3c092a), so a merge would only invalidate every byte-identical measurement above. Recorded rather than silently absorbed.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_012zGPuVVX3deAx9LdjK8jCk

---

Generated by Claude Code