test(scripts): batch 6b — assertion floors for nine scripts/** self-tests (sink repair + hoisted battery) - #15248

Part of #13799 (batch 6b — the class-1 + class-2 deferrals from PR #15003's table: sink repair plus a single hoisted battery).

Dispatched by PM claim 5536969005; dev claim 5536991755. Branched at origin/main = 1bc3c092a, the dispatch pin, re-read and still matching at the branch point. origin/main was merged in at 332ada04c (a real merge commit — no rebase, no force-push) after PR #15217 and three siblings landed; the merge touches neither of the two files this PR edits.

2 of the 9 dispatched files are floored here. 7 are deferred with a measured reason rather than improvised, per the dispatch's own instruction to defer table-driven / bare-throw files to the loop-body decision. The deferral is the finding of this batch, and the evidence is below.

Floored (2)

Both declare SELF_TEST_BATTERY_FLOOR = 1 (the roster's own size, so deleting the entry cannot silence the floor), UNATTRIBUTED_BATTERY, and a floor evaluated before the verdict that refuses a below-floor or DID-NOT-RUN battery and an undeclared one. Neither file carries two named section banners, so each declares ONE battery opened at the top of the self-test body — the PR #14896 / #15003 / #15217 hoisted shape. No comment is promoted to a section head. No assertion condition is inverted, rewritten, reordered or re-judged; the case objects are passed through untouched and the loop that runs them is not touched at all. A floor breach is filed into each file's own existing failure sink (cases with ok: false; the failed counter), so the existing verdict reds on it with no verdict line rewritten, and the #13798 handshake is byte-unchanged.

Case count before == after, measured on runs

Line-level identity of the sink rewrite

Removed lines that do not reappear in the added lines after normalising indentation:

• check-pnpm-acquisition — 1 residue, const t = (name, ok, detail) => cases.push({ name, ok: Boolean(ok), detail });. Fully explained: the concise arrow was split into head + registerCase(); + body, and its expression body cases.push({ name, ok: Boolean(ok), detail }); reappears verbatim inside the block body.
• check-skill-frame-freshness — 12 residues, all of them the opener line cases.push({ rewritten to addCase({. The case object literals that follow never enter the diff at all, which is the point: nothing about any case changed.

Ablation (each leg proves its mutation landed on disk; each restore is proven by HEAD-blob hash equality, not by an exit code)

Both files are plain node scripts/NAME.mjs entrypoints — executed from source, never resolved through a package exports into a dist/, so no rebuild leg applies and no stale-artefact false green is available here. Both ablations ran against a committed implementation, so the restore leg had a real reference point. The harness carries trap ... EXIT INT TERM with absolute paths.

One ablation leg was rejected by its own landing check before it could produce a reading: the first registerCase()-removal attempt passed an anchor with a trailing newline, so grep -c read it as two patterns (the empty one matching every line) and reported 685 -> 684. The run stopped, the trap restored the tree, and the leg was re-run with a corrected anchor. Recorded because the first attempt produced no measurement, not a passing one.

extractWatchHints on both trees, per file

Run through the real extractWatchHints exported by scripts/pm/dispatch-gates.mjs (untouched by this PR), on the pin's content and this branch's content:

check:watch-hint-literal green: 49 declarations across 4 rostered names.

The card's instrument, with the concurrent batch accounted for rather than asserted

node scripts/measure-self-test-floor.mjs --json, at the pin and on this branch:

11 files moved NONE to ROSTER across that range. 9 are PR #15217 (batch 5), which merged in between and is in this branch through the merge commit; 2 are this PR's. Unexplained movement: none. Symmetric difference of (moved, minus batch 5's nine) against this batch's floored worklist: empty. So the NONE drop attributable to this PR is exactly 2, the number floored.

All 7 deferred files read exactly as they did at the pin — six NONE, one COUNT — i.e. untouched.

Deferred (7), each with the measurement that decided it

⛔ Not improvised. The dispatch reserves table-driven and bare-throw files for the loop-body decision, and all seven measure as the same shape: a literal const cases = [ ... ] table, zero cases.push sites, exactly one loop over the table, and sink writes that are failure-only.

Why the recipe cannot be applied to these without the reserved decision. PR #15003's table counts the failed += 1 / failures++ sites as "inline sites", but every one of them is lexically inside a failure branch or a catch — verified on all seven, guard by guard. So routing them through a block-bodied helper that calls registerCase() would register a case only when a case fails: a fully green run would register zero cases and the battery would read DID NOT RUN. That is the floor inverted, not installed — the one outcome this card exists to prevent. There is no per-case sink to repair, and the only site that executes once per case regardless of outcome is the loop body itself, which is exactly the loop-body decision the dispatch reserved. Batch 4's thunk is not an escape either: in 4 of the 7 the loop body carries continue, which cannot cross a function boundary, so check(() => { ... }) cannot wrap those bodies verbatim.

This matches the card's own Tier C wording — "table-driven self-tests where the natural roster is the table's own named rows rather than sections" — so the natural remedy for these seven is a roster of the table's named rows, which is a shape no batch has landed yet.

One re-verification correction to carry forward: check-refd-timer-probe.mjs reads COUNT, not NONE, on the instrument — it is one of the two files on the tree that derive and print a case count nothing compares (the other is pm/check-half-states.mjs, excluded by the dispatch). The card body names a different pair, which the two intervening batches have since moved.

Verification

Exit codes captured by redirect before any pipe; every result below is the gate's own printed verdict line, never a bare status variable. All runs in the foreground, on final commit 332ada04c (git rev-parse --short HEAD from the same tree the runs used).

Derived family — node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, which derives its own change set from the merge base (no hand-passed path list). The first derivation refused to be trusted, printing STALE TREE and naming check-ratchet-remedy-authority.mjs and dispatch-gates.mjs as changed under it; origin/main was fetched and merged and the family re-derived clean. 23 commands, all exit 0.

Always-runs tail: check:pm-dispatch-gates (1353 cases pass) · check:self-test-wired (168 scripts, every one CI runs) · check-self-test-workflow-commands and its --self-test · check:declared-population-live (201 of 252 families reach this tree's 8281 files) · check:ratchet-remedy-authority (202 scripts swept) · check:watch-hint-literal · check:nul-bytes (8274 files scanned, no raw control bytes). All exit 0.

Each edited script's own suite, normal mode against the tree and --self-test, all exit 0 — both are lint.yml gates and both stay green on the tree:

• check-pnpm-acquisition normal + --self-test (12 cases pass)
• check-skill-frame-freshness normal + --self-test (12 cases pass); normal mode reports the decision frame current with origin/main, 2 copies across 2 files

Neither script is named by any test file in the tree (git grep over test globs: no hits; referenced only from .github/workflows/lint.yml and package.json), so no separate suite is owed.

Control-byte scan of both edited files and of every body posted to GitHub: clean.

skip-changeset: this PR publishes nothing from any package — it edits two scripts/** gate self-tests. Label applied through the additive endpoint and read back.

Generated by Claude Code