tooling(scripts): row batteries for four table-driven self-tests (#13799 batch 7b)

[claude]

Part of #13799

Batch 7b of the assertion-floor card: four table-driven scripts/** self-tests, under recipe A exactly as the batch 7 pilot (PR #15271, ACCEPT 5537827823) settled it. Every file had failures === 0 as its ONLY success condition, so "every case held" and "the cases never ran" printed the same line. Four files floored, nothing deferred.

All evidence below was produced at the final commit 5dfa45644, against an installed base worktree at the dispatch pin origin/main = 4dd5041bd.

Rows to batteries to floors

file	rows	batteries declared	floor / row	roster size pin	assertions NOT floored (stated in place)
scripts/pm/check-skill-id-lint.mjs	14	14	1	14	none — every assertion is a table row
scripts/check-agent-model-declared.mjs	18	18	1	18	the 4 discovery-block assertions below the loop (verdict has always printed 18 rows, not the 22 assertions that run)
scripts/check-startup-registry-verdict.mjs	18	18	1	18	the 8 expectRoot(...) assertions of the dead-root reverse proof
scripts/js-comment-mask.mjs	23	23	1	23	the 12 interpolation-view assertions pushed onto extra at runtime

Recipe A, per file: battery name = the row label verbatim, no prefix; registerCase(label) is the FIRST statement of the driving loop body; the roster is a LITERAL the table is checked against, never derived from it; floors, not equalities; the roster's own size pinned at the row count; the red-run line becomes N failure(s) (cases and floor) and the green line is byte-identical to main's.

check-agent-model-declared's driving loop carries three continues (one in a catch), so registration sits ahead of the try — per Q1 of the pilot ACCEPT, the floor asserts REACH, and a registration behind a guard reintroduces the very inversion the recipe exists to avoid.

Two "not a row" scope calls, both written into the file. They are not deferrals — the tables themselves are fully floored:

• check-agent-model-declared / check-startup-registry-verdict: those assertions are written inline one at a time below the loop, and each file's verdict line has always counted the table's rows separately from them.
• js-comment-mask: the interpolation section is not a second literal table. Its rows exist only because an x(...) call pushed them onto extra at runtime, so a roster taken from that loop would be DERIVED — a deleted call would silently delete its own floor, the one defect this shape exists to prevent. Flooring it needs a literal roster of its own and is not a table-row question.

Roster ⇄ table cross-check

Roster keys printed against the labels the loop actually registered, per file: 14 / 18 / 18 / 23, symmetric difference EMPTY in all four. Both halves of the duplicate refusal are in place: a duplicate ROSTER key collapses and trips the size pin, a duplicate TABLE label is named by the cross-check.

Ablations — 16 legs, 4 per file, all from the committed implementation

leg	result (all four files)
one row deleted	exit 1, names that row DID NOT RUN (roster keeps the label; 2 to 1 occurrences proved on disk)
one row renamed in the table only	exit 1, names both: the undeclared registration AND the declared name that DID NOT RUN
registerCase() call removed	exit 1, every row DID NOT RUN — 14 / 18 / 18 / 23 lines — with all original assertions still green (14 / 22 / 26 / 35 ticks)
size pin set unreachable (999)	exit 1, breach line names the count: declares 14 / 18 / 18 / 23 batteries

The third leg is where the row count is measured on runs rather than read off the source, and it agrees with the roster size in every file.

No rebuild leg is owed: these are plain .mjs scripts executed directly from source, with no dist on the resolution path. Every mutation was proved on disk before the run (occurrence counts of the removed and injected text, plus byte deltas — never an editor's exit code), and every restore is proved by HEAD-blob hash equality plus an empty git diff HEAD; each leg ran under a trap ... EXIT INT TERM with absolute paths.

Both modes vs the base worktree

For all four files, --self-test and normal mode: stdout and stderr byte-identical to the base tree's, exit code identical (0 on both; js-comment-mask's no-argument usage path is 2 on both).

js-comment-mask is a library — the importer proof

Six importers, --self-test on both trees: check-comment-mask-corpus, check-comment-mask-adoption, check-error-code-casing, check-entry-guard, check-declaration-mirrors, check-watch-hint-literal — all exit 0, stdout and stderr byte-identical base vs branch. The same six also run green in normal mode on the edited tree, which is the real consumption path.

Census and hints

node scripts/measure-self-test-floor.mjs --json: NONE 35 to 31, ROSTER 132 to 136, COUNT unchanged at 2. The movers are exactly the four files of this batch — symmetric difference against the worklist EMPTY, and no file newly reads NONE.

extractWatchHints output is identical on both trees for all four files, and check:watch-hint-literal is green.

Gates

The derived family for this diff — node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, 30 commands, derived at 5dfa45644 — plus the always-runs tail. All green, exit codes captured by redirect before any pipe.

Two of the thirty first came back NOT MEASURED on a fresh worktree rather than red — check:doc-formula-expressions (exit 3) and check:browser-reachable-entries — both refusing because the packages they read compiled output from were not built. Neither is a finding and neither can be reached by this diff; both were re-run green after building @objectstack/spec, @objectstack/formula and @objectstack/lint through the shared verify lock. Also green: check:pm-dispatch-gates, check-self-test-wired (both modes), check-self-test-workflow-commands (both modes), check:declared-population-live, check:ratchet-remedy-authority, check:watch-hint-literal, check:nul-bytes.

Control-byte scan clean over every edited file and every posted body.

Invariants

⛔ No assertion condition touched. ⛔ No control flow rewritten. ⛔ No comment promoted to a section head. The verdict handshake is untouched in each file, in each file's own spelling (sentinel return in check-skill-id-lint and js-comment-mask, the selfTestReachedVerdict flag in check-agent-model-declared, flag plus exit code in check-startup-registry-verdict). The only four deleted lines in the whole diff are the four red-run verdict lines; everything else is additive.

skip-changeset: no package publishes anything from this diff.

---

Generated by Claude Code

---

Generated by Claude Code

[claude]

Out-of-scope finding filed while landing this batch, unassigned and finding-labelled, not queued: #15296 — recipe A floors the table rows only, so 24 inline assertions in three of these four files (4 discovery + 8 expectRoot + 12 interpolation) still register nothing on a green run, while the census now reads those files as ROSTER. None of those 24 assertions was touched here; each scope call is written into the file it belongs to.

---

Generated by Claude Code