test(scripts): give six more self-tests a battery roster and floor (#13799 batch 3)

[claude]

Part of #13799 (batch 3, the class-1-ONLY deferrals from PR #15003's table).

PR #13487's roster-and-floor shape, exactly as batches 1/1b/1c/2 landed it (PRs #14851, #14872, #14896, #15003), transplanted onto the files whose only recorded defect in #15003's table was the assertion-sink shape while the banner count and the body-level verdict site already qualified. 39 batteries / 306 cases pinned across 6 files. The seventh candidate is deferred with a measured reason rather than forced.

Per file, and nothing else:

• a module-level frozen SELF_TEST_BATTERIES roster of names with a per-battery floor, SELF_TEST_BATTERY_FLOOR pinning the roster's own size, and UNATTRIBUTED_BATTERY, placed after the import block;
• battery('BANNER TEXT') at each existing section banner, named verbatim from the banner with its marker stripped;
• one registerCase() at the head of the one in-body assertion helper this PR gives each file — which is the whole of the class-1 work;
• the floor evaluated immediately before the existing verdict site, its messages entering the file's own failure sink so the report the file already prints carries them.

⛔ No case is rewritten, none reordered, no assertion changes meaning. ⛔ No TOTAL is pinned.

What "class-1" cost per file, since that was the deferral reason

The three sink shapes in #15003's table needed three different — and each strictly local — repairs:

shape	files	what this PR did
module-level assert	check-adr-links, check-doc-anchors, check-kernel-hook-pairs	a thin in-body wrapper check(cond, message) that registers the case and then defers to the existing assert, whose semantics (print, process.exit(1) on the first failure) are untouched. Call sites change by one token: assert( becomes check(.
sink written at inline sites	check-published-files (7 sites), ablation-dist-preflight (6 sites)	one block-bodied helper replacing the inline writes. Message text is byte-identical; the only behavioural change is that the case is COUNTED whether it holds or not, which is exactly what lets the floor tell "held" from "never ran".
concise-arrow sink t	pm/check-clause2-carriers	a block body for t, so it can hold registerCase().

In the inline-sink pair the failure-message template is now evaluated eagerly rather than only on the failing branch. Every one of those templates interpolates values already computed on the line above, so this is free of side effects; it is recorded because it is the one place where the diff is not a pure re-routing.

The floor, and case count before == after

Both numbers are measured on a run, not derived from the diff. before for check-clause2-carriers is the count its own verdict line already printed; for ablation-dist-preflight it is the ✓ lines the base run prints, one per case; for the other four the base file was instrumented with a counter at the point each case is evaluated, run, and then restored — restore proven by git hash-object equalling the HEAD blob and by an empty git diff HEAD, not by the exit code of the restore command.

file	batteries	cases before	cases after	how before was read
scripts/ablation-dist-preflight.mjs	3 (first hoisted)	36	36	✓ lines on the base run
scripts/check-adr-links.mjs	4	23	23	instrumented base assert
scripts/check-doc-anchors.mjs	8	22	22	instrumented base assert
scripts/check-kernel-hook-pairs.mjs	10	16	16	instrumented base assert
scripts/check-published-files.mjs	2 (first hoisted)	59	59	counter at each of the 7 evaluation sites
scripts/pm/check-clause2-carriers.mjs	12	150	150	the verdict line's own count
total	39	306	306

Two files run cases before their first banner (ablation-dist-preflight, check-published-files); their first battery is opened at the top of the body and that banner then carries no second opener — the PR #13487 precedent's own shape, as batches 1b and 2 did it.

check-published-files's 59 reconciles against its own tables independently of the instrument: 21 pattern + 12 classification + 4 population-declaration for the first battery, and 12 exports verdict + 6 (the Fix:-line loop skips the 6 rows expected to pass) + 3 census-floor + 1 shared enumerator for the second.

The card's own instrument agrees

node scripts/measure-self-test-floor.mjs, which the card names as the thing to re-measure with rather than re-derive:

	base ebb082204	this branch
unfloored (hole 1)	61 of 167	55 of 167
floored by a roster	104	110

+6, and the six new roster entries are exactly this PR's six files.

Ablation: each floor was made to fail, on the battery it names

Per file: the first declared battery's opener is renamed so that battery registers nothing, the self-test is run, and it must exit non-zero naming that battery. Then restore, and it must be green again. Both legs are proven on disk — the mutation by counting the injected marker, the restore by git hash-object against the HEAD blob plus an empty git diff HEAD and a zero residual-marker count — never by an editor's or a git checkout's exit code, and every script carries a trap ... EXIT INT TERM restore.

These are .mjs scripts executed directly from source: no dist/ and no exports resolution stands between the edit and the run, so the rebuild leg the ablation contract requires for package-resolved subjects does not apply here. Recorded rather than skipped silently.

file	ablated battery	ablated exit	named it	restored exit
ablation-dist-preflight.mjs	whole-tree accounting: the pure table	1	✓ DID NOT RUN	0
check-adr-links.mjs	Discrimination: verbatim link shapes are invisible; prose ones are not.	1	✓ DID NOT RUN	0
check-doc-anchors.mjs	Slug parity with the renderer, on the two shapes that actually broke.	1	✓ DID NOT RUN	0
check-kernel-hook-pairs.mjs	A hook pinned on both sides passes.	1	✓ DID NOT RUN	0
check-published-files.mjs	the dispatch-gates declaration (#10542)	1	✓ DID NOT RUN	0
pm/check-clause2-carriers.mjs	the declaration reader: the fixed spelling, and everything that is not	1	✓ DID NOT RUN	0

⭐ The one behavioural hazard #15003 measured, measured again here

#15003 deferred scripts/pm/dispatch-gates.mjs because its self-test reads its own source and four battery names carrying path-shaped text were read by its own path-population extractor as declared populations — a change of meaning, not a no-op.

This batch adds new string literals to six files that dispatch-gates reads as gate sources, so the same extractor could have read a battery name as a watch hint. It does not. extractWatchHints was run over all six files on both trees:

file	hints before	hints after
ablation-dist-preflight.mjs	[]	[]
check-adr-links.mjs	["docs/adr"]	["docs/adr"]
check-doc-anchors.mjs	["content/**","README.md/**","ARCHITECTURE.md/**"]	identical
check-kernel-hook-pairs.mjs	["packages/core/src"]	["packages/core/src"]
check-published-files.mjs	13 hints	identical, same order
pm/check-clause2-carriers.mjs	[]	[]

Byte-identical on every file. check:pm-dispatch-gates (1343 cases) and check:declared-population-live are both green on the branch, which is the same finding reached the other way.

Deferred (1), with the measurement

file

measured reason

scripts/partition-test-shards.mjs

shape ✓ on two of three, sink ✗ — and the class-1 repair here is not a re-routing. 7 named banners and a body-level verdict site, as #15003 recorded. But its 36 assertions are bare throw at every site, and only 22 are the uniform single-line if (cond) throw new Error(msg);. The other 14 are multi-line, several nested inside for loops and closures. Routing them through a helper means inverting the failure condition by hand at each site — and a mis-scoped or dropped ! yields an assertion that still registers its case, still prints nothing, and still passes, so the floor being installed cannot see it. That is a rewrite of every assertion rather than the "route without changing assertion semantics" the shape asks for, so it is deferred rather than improvised past. The route a later batch can take without inverting anything: a thunk helper, const check = (fn) => { registerCase(); fn(); };, wrapping each site's existing if (...) throw ... verbatim — no condition is touched, at the cost of reindenting the 14 multi-line blocks.

⛔ Untouched by construction, per the dispatch: scripts/pm/dispatch-gates.mjs (still DEFERRED, reason above), scripts/pm/check-half-states.mjs (~17 kloc, the analyzer cannot resolve its body), every class-2 / class-3 file in #15003's table, and the files held by other live cards.

Shrink-only ledgers: scanned, one worth naming

Every battery whose cases come off a module-level list was enumerated and each list read, per the card's "a floor is not always the measured count":

• CONVENTION_ILLUSTRATIVE_TARGETS (check-adr-links, 3 rows feeding 6 of battery 3's 8 cases) is a pin of three strings that ADR-0046 must keep containing — the file's own header says so. It grows or holds; it is not a baseline that legitimate work shortens. Named here because it is the one floor in this batch that moves with a list at all.
• KNOWN_DEAD_TARGETS (check-adr-links) is shrink-only, and is exactly the case the card warns about — so note what it drives: one structural case (no entry has outlived its excuse), not one per row. It is currently [], and shrinking it cannot move that count. This is the card's own remedy — pin the part that does not move with the list — arrived at without needing to deviate.
• ROOT_DIR_WATCH_HINTS, FORBIDDEN, EXPORTS_CENSUS_FLOOR (check-published-files) and CLAUSE2_VALUES (check-clause2-carriers) are registers and closed sets that grow or are pinned by an adjacent assertion; none is shrink-only.

Verification

Every exit code captured by redirect before any pipe; each gate's own verdict line read rather than a bare $?. Run on 6bddfc959, the final commit of this branch.

• Family derived from the tree, not from a hand-written diff: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack → 27 commands, all 27 run, all exit 0.
• The always-runs tail: check:pm-dispatch-gates (exit 0, 5m39s, 1343 cases), check-self-test-wired + --self-test, check-self-test-workflow-commands + --self-test, check:declared-population-live, check:ratchet-remedy-authority, check:nul-bytes — all exit 0.
• Each of the six in both modes: its --self-test and its normal gate run against the live tree. check-adr-links (599 destinations resolve) and check-doc-anchors (298 fragment links across 409 sources) are lint.yml gates and are green on the tree. ablation-dist-preflight is a per-invocation tool, not a repo sweep: with no arguments it prints usage and exits 2, unchanged from the branch point.
• Neither of the two files that name an edited script (packages/qa/dogfood/test/multi-package-artifact.dogfood.test.ts, packages/plugins/plugin-auth/src/rate-limit-storage-isolation.test.ts) exercises it — both are prose references in comments. Each script's --self-test is its own suite, and all six ran.
• Control-byte scan over all six edited files (grep -naP '[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]'): no match. check:nul-bytes green (8238 files scanned).

skip-changeset: scripts/** publishes from no package, the same call PR #15003 made for the identical surface.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_012zGPuVVX3deAx9LdjK8jCk

---

Generated by Claude Code