test(scripts): batch 8b — assertion floors for two b4-thunk and two recipe-A self-tests

[claude]

Part of #13799

Batch 8b of the self-test assertion-floor card: two b4-thunk files and two recipe-A files, from the batch 8 census table (5538556734) as dispatched by its ACCEPT (5538595066) and the PM claim 5538595337. Base pin origin/main = 5b09356b7; every reading below was taken at head 9965c88d3. Exactly four files, skip-changeset, no deferrals.

Per file

file	census row	recipe	roster	floor	measured on a run
scripts/measure-test-shard-timings.mjs	16	b4 thunk	1 hoisted battery, size pin 1	22	registered 22 case(s), below its pinned floor of 23
scripts/measure-position-name-fold-census.mjs	15	b4 thunk	1 hoisted battery, size pin 1	24	registered 24 case(s), below its pinned floor of 999
scripts/check-override-consistency.mjs	7	recipe A	17 row labels, floor 1 each, size pin 17	17	declares 17 batteries, below the pinned 999
scripts/pm/check-skill-line-ratchet.mjs	19	recipe A	155 row labels, floor 1 each, size pin 155	155	declares 155 batteries, below the pinned 999

A measured correction to census row 19

The row warned that the cases array "also spreads an IIFE-produced block — generated rows cannot take a literal roster key", and dispatched the file as "recipe A on the literal rows only, generated rows outside the roster".

Measured, that premise does not hold for this file. Each of the four ...(() => { ... })() spreads is a SCOPING device: it declares local fixture consts and then return [...]s an array of LITERAL [label, actual, expected] rows. No row label is a template string, none is computed, and no row is produced by a map / push / loop. Three independent readings agree on 155:

• 155 labels extracted from source by indentation (21 at the array's top level, 134 inside the four IIFEs), all single-quoted literals, zero duplicates;
• 155 literal row starts, and only 4 spreads, all top-level;
• 155 cases pass. printed by the green line on a run of the unmodified base tree.

So all 155 rows are floored, with no residue — nothing here is the extra-call class of PR #15286, and partitioning the table would have been the lossy reading (the 7a Q1 precedent). The reasoning is recorded in place above the roster.

Where the floor lives in measure-position-name-fold-census

main() gates the PRODUCTION census on selfTest({ quiet: true }), so this file has two callers of one body. The floor is evaluated inside the !quiet branch: the quiet call prints no verdict and therefore makes no claim for a floor to guard, and evaluating it there would let a roster edit change what a census run outputs. Proven, not argued — leg D4 below.

One site stays outside the roster and says so in place: the catch arm of the AUDIT_CONTROLS loop ends in continue, which is illegal inside the thunk's arrow function, and rewriting that control flow is exactly what the verbatim rule forbids. Its sibling in the same loop body, if (!ac.expect(result)), is floored, so a loop that stops running still reds.

Line-level identity (thunk files)

Every removed line, stripped of leading whitespace, reappears among the added lines. No condition was inverted; the dropped-! hazard cannot arise.

file	removed	added	removed-not-in-added
measure-test-shard-timings.mjs	38	191	0
measure-position-name-fold-census.mjs	52	193	0

Wrapped-site count == assertion-site count in both: 22 check(() => { wrappers over 22 if (...) throw ... sites; 13 wrappers over the 13 if (...) problems.push(...) statement-sites carrying 14 of the 15 push sites (one if/else if carries two), the 15th being the declared catch-arm exclusion.

The two recipe-A files delete only sink lines: check-skill-line-ratchet deletes exactly one, its red-run verdict, reworded to the recipe's N failure(s) (cases and floor); check-override-consistency deletes four, converting its boolean passed into a failures counter so the same red line can carry a count. No assertion condition is touched in either, and both green lines are byte-identical to the base tree's.

Roster ⇄ table cross-check (recipe A)

Both lists printed and compared; symmetric difference empty in both directions, order identical, zero duplicate labels.

• check-override-consistency: 17 roster keys vs 17 table labels.
• check-skill-line-ratchet: 155 roster keys vs 155 table labels.

Ablation ledger — 15 legs

Every leg was run from the committed implementation. Each proves its mutation landed on disk (a grep count of the removed anchor AND of the injected text; a no-op mutation aborts the leg and discards the reading), and each restore is proven by HEAD-blob hash equality plus an empty git diff HEAD — an empty hash is read as failure, never as "nothing to compare". The mutators carry a trap ... EXIT INT TERM restore against an absolute REPO_ROOT.

leg	file	result
A1 row deleted	ratchet	names "over the ceiling -> red" DID NOT RUN, exit 1, 0 failing cases
A2 row renamed	ratchet	both halves: the undeclared registration and the missing declared name
A3 registerCase() removed	ratchet	all 155 DID NOT RUN, every case still green
A4 size pin unreachable	ratchet	declares 155 batteries, below the pinned 999
B1 row deleted	override-consistency	names the row, exit 1, 0 failing cases
B2 row renamed	override-consistency	both halves
B3 registerCase() removed	override-consistency	all 17 DID NOT RUN, every case still green
B4 size pin unreachable	override-consistency	declares 17 batteries, below the pinned 999
B5 duplicate label in the TABLE	override-consistency	names the colliding label, and the shadowed row as DID NOT RUN
C1 battery opener renamed	shard-timings	undeclared battery registered 22 + declared one DID NOT RUN
C2 registerCase() removed from the thunk	shard-timings	DID NOT RUN — 0 cases registered, 22 pinned, cases green
C3 roster size pin unreachable	shard-timings	declares 1 batteries, below the pinned 9
D1 battery opener renamed	fold-census	undeclared battery registered 24 + declared one DID NOT RUN
D2 registerCase() removed from the thunk	fold-census	DID NOT RUN — 0 cases registered, 24 pinned, cases green
D3 roster size pin unreachable	fold-census	declares 1 batteries, below the pinned 9
D4 floor made unreachable, PRODUCTION path run	fold-census	--self-test reds (exit 1) while the census run stays exit 0 and byte-identical to the base tree — the floor cannot reach production

These scripts are executed directly by path, not resolved through a package exports, so no dist/ rebuild leg applies; there is nothing between the edited source and the run.

Both modes vs an installed base worktree at the pin

A detached worktree at 5b09356b7, pnpm install exit 0. Every stream compared with cmp.

file	mode	base exit	branch exit	stdout	stderr
fold-census	normal	0	0	identical, 4225 B	identical, 0 B
fold-census	--json	0	0	identical, 6352 B	identical, 0 B
fold-census	--self-test	0	0	identical, 430 B	identical
shard-timings	normal (usage)	1	1	identical, 0 B	identical, 75 B
shard-timings	--self-test	0	0	identical, 41 B	identical
override-consistency	normal	0	0	identical, 1627 B	identical, 0 B
override-consistency	--self-test	0	0	identical, 1340 B	identical
skill-line-ratchet	normal (reads skills/**, .claude/**)	0	0	identical, 7917 B	identical, 0 B
skill-line-ratchet	--self-test	0	0	identical, 9288 B	identical

18/18 streams identical. skills/** was never edited; the ratchet's normal mode is green on both trees.

The card's instrument

node scripts/measure-self-test-floor.mjs --json on both trees, 169 rows each:

• NONE 34 → 30 — drops by exactly 4;
• ROSTER 133 → 137; COUNT 2 → 2, unchanged;
• the movers are exactly the four worklist files, symmetric difference against the worklist empty in both directions.

extractWatchHints (from scripts/pm/dispatch-gates.mjs) over all four files, base vs branch: byte-identical ×4 — the 172 new roster labels contribute zero hints, and check-skill-line-ratchet's own 28 hints are unmoved. check:watch-hint-literal green.

Gates (foreground; exit codes captured by redirect BEFORE any pipe)

• Derived family: 28/28 green, from node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack (change set: exactly the 4 files vs merge base 5b09356b7). check:pm-dispatch-gates under the shared verify lock, VERDICT command-exit 0 · held the lock 374s, 1381 cases pass.
• Always-runs tail: check-self-test-workflow-commands.mjs and its --self-test, check:declared-population-live, check:ratchet-remedy-authority, check:watch-hint-literal, check:nul-bytes, check-self-test-wired.mjs and its --self-test — all exit 0. NOT MEASURED: pnpm check:self-test-wired exits 254 because no such package.json script exists (0 occurrences); the wired spelling is the direct node invocation, which is green.
• Each edited script's own suite: all four run in normal mode AND --self-test; the only non-zero is measure-test-shard-timings.mjs normal mode, exit 1 = its usage line, identical to the base tree's.
• Pin suite naming an edited script: packages/plugins/plugin-security/src/position-name-fold-warning.test.ts — 28/28 pass, after building the dependency closure under the verify lock.
• Repo-wide pnpm lint exit 0, not narrowed.
• Control-byte scan (grep -naP '[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]') clean on all four files and on every body posted.

Derivation carried a STALE TREE warning: origin/main has moved to ed217e697 (2 commits) since the dispatch pin, and 9 files changed across that range — all of them 7a/7b's landed files, disjoint from these four. The branch is deliberately left on its pin rather than rebased.

🤖 Generated with Claude Code

https://claude.ai/code/session_012zGPuVVX3deAx9LdjK8jCk

---

Generated by Claude Code