fix(cli-gate): name it.skip / it.only / it.each blocks instead of collapsing them onto (top-level)

[os-litant]

Fixes#12545

callbackSiteName() required a bare-identifier callee, so it('x', fn) was named but it.skip('x', fn) (a PropertyAccessExpression) and it.each(table)('x', fn) (a CallExpression) returned null. enclosingFunctionName() then fell out of its loop to '(top-level)', collapsing every skipped / only / table-driven block in one file onto a single site key — the same collision #12531 closed for the identifier branch, one modifier over.

Measured on origin/main @ aa4591971 before touching anything:

call	callee node	before	after
it('x', fn)	Identifier	it("x")	it("x") (unmoved)
it.skip('x', fn)	PropertyAccess	(top-level)	it.skip("x")
it.only('x', fn)	PropertyAccess	(top-level)	it.only("x")
describe.skip('x', fn)	PropertyAccess	(top-level)	describe.skip("x")
it.each([…])('x', fn)	CallExpression	(top-level)	it.each("x")
describe.each([…])('x', fn)	CallExpression	(top-level)	describe.each("x")
it.skip.each([…])('x', fn)	CallExpression	(top-level)	it.skip.each("x")
promise.then(cb)	PropertyAccess	(top-level)	(top-level) (refused)
rows.map(cb)	PropertyAccess	(top-level)	(top-level) (refused)

The unwrap is targeted, and it is gated on ROSTERS rather than on node kind

⛔ A blanket "property access whose base is an identifier" rule was measured, not assumed, and it is much worse than the defect. Against the gate's own live population (packages/cli/test/**, 98 files), that rule matches:

• 209 call sites across 87 distinct callees — literals.filter, child.on, entries.map, warnings.filter, server.close, Array.from, vi.fn, run.then, rows.map …
• against this roster's 22 — it.each(…) ×20 and describe.each(…) ×2, which is the entire modifier population of that directory.

Those 87 are not test blocks, their callee is a worse site name than the enclosing block the walk already reaches, and naming them would move the key of every site nested inside one. Re-keying is a data migration wearing a bug fix's clothes.

So the gate is two rosters, TEST_BLOCK_ROOTS × TEST_BLOCK_MODIFIERS:

• roots — it, test, describe. ⭐ A roster on this side is load-bearing, not tidiness: this directory calls it.name, it.type, it.next and it.prevVersion on ordinary loop variables that happen to be spelled it. Gating on the root alone would name a callback site after one of those.
• modifiers — skip / only / todo / fails, concurrent / sequential / shuffle, each / for, skipIf / runIf. Every segment after the root must be on the roster, so it.skip.each(…) is named through the whole chain and a.b.c(cb) never is. each/for and skipIf/runIf curry, and are unwrapped exactly one call deeper.
• ⛔ extend is deliberately absent: test.extend(fixtures) builds a new test function, not a block, and its product is bound to a name the identifier branch already reads.

The title is read from the outer call's arguments, so it.each(['from the table'])('from the block', fn) is named for the block, never for a string sitting in the table. That is pinned.

⭐ Key stability — the proof, stated rather than left to be noticed

The identifier branch is reached first and unchanged, so plain it(…) naming is byte-identical. That matters because DELIBERATE_REROUTE holds two keys this function produced, carrying it("…") titles verbatim, and DELIBERATE holds two more from the variable-declaration branch — all four pinned in both directions, so a moved key exits 1.

⇒ The green live-tree run below IS the proof that no existing key moved, and it says so in its own words:

✓ check:cli-test-child-env: 35 spawner source(s) among 98 under packages/cli/test/**; …
(0 baselined in 0 file(s), ⛔ SHRINK-ONLY; 2 deliberate site(s) and 2 deliberate reroute(s) still pinned).

And the pin was proven LOUD rather than assumed to be, because a green that cannot fail proves nothing. Reverse verification on the committed tree — predicted direction RED:

• mutated one DELIBERATE_REROUTE key (NODE_ENV=development (explicit) → NODE_ENV=deliberately-moved-key (explicit));
• mutation confirmed on disk, not by the editor's exit code: anchor occurrences 1 → 0, mutant 0 → 1, blob 2210188b… → ec97c08b…;
• gate on the mutated tree: exit 1, reporting the now-unclassified site by the very name this function produces — [it("NODE_ENV=development (explicit): the gate stays OPEN — unaffected by the production default")];
• restored with git checkout HEAD -- naming an ABSOLUTE path, and the restore was proven by state: blob back to 2210188b…, git diff HEAD empty.

⛔ No registry entry was edited. Had one gone stale, that would have been the finding.

Self-test: 132 → 152 cases, extend-never-weaken

Exactly one pre-existing case changed, and it is the one that pinned the defect being fixed:

'an it.skip() callee is a property access, not an identifier, so it still reads (top-level)' → 'an it.skip() block is named, not (top-level)'

That is a strengthening — a specific name replaces a fallback — and what the old case's comment was really protecting (the refusal of promise.then / rows.map) is now pinned outright, as six explicit refusal cases rather than as a side effect. Every other change is additive; the full deletion list in the diff is that one case plus the two implementation lines and the docblock limit it described.

Added: each modifier form; both curried forms; the chained form; the table-string trap; two sibling it.each blocks getting distinct names; an end-to-end audit() case where one registry entry silences one it.each block and leaves its sibling a finding; the six refusals; and —

⭐ the negative pin, which is the one that matters: 'a plain it() block STILL yields exactly it("x") — the identifier branch did not move'. Without it nothing in the suite distinguishes "the modifiers now work" from "everything got renamed".

Rule 5 — the second implementation, MEASURED, not swept

scripts/check-durability-degradation-log-level.mjs:1695enclosingFunctionName(node), used at :2451 and :2493. ⛔ Not touched here. What I read:

It does NOT have this card's defect — and not because it handles modifiers, but because it has no callback-naming branch at all. Its four shapes are function declaration, method declaration, variable-bound arrow/function-expression, and named function expression. It never names it('x', fn)either, so there is no identifier branch for a modifier unwrap to extend. The claim comment's guess that "it takes no sourceFile argument, so it may not have the same branches" is correct.

Its fallback is undefined, not '(top-level)', and the two call sites diverge:

• :2493declaredPropagationFor — no defect, structurally.if (!fnName) return globalPropagation; short-circuits, so no key is ever built. That is a closed door, not a shared key.
• :2451 → readInventionKey (:2827) — the same collision class, latent. It keys a shrink-only baseline on f.file and f.fn joined by ::, substituting a fixed placeholder — the word anonymous wrapped in angle brackets — whenever f.fn is undefined. So two anonymous-callback read-seams in one file share one key, and a single baseline entry would classify both.

Measured live:scripts/durability-read-invention.baseline.json holds 1 entry, with 0 anonymous fn and 0 files carrying more than one entry. ⇒ latent, exactly as this card was. Filed unlabelled as its own card: #12576.

Falsifications of the dispatch order (reported, not reconciled)

• ⚠️main had moved. The order pins the merged base at 1b7e3d2ce; origin/main was aa4591971 when I cut the branch (and 9bed0b0fe by the time gates ran). Re-verified the defect on aa4591971 rather than trusting the description — it reproduced exactly.
• ⚠️The card's "24 such spellings" is a textual count including 2 prose mentions.grep finds 24 under packages/cli/test/**, but two are inside docblocks (it.each` below against silently iterating nothing, it.each` over an empty). The AST count is 22. Fittingly, the suite's own section (6) exists to make exactly this point — "prose is not code, which is what an AST buys over a text scan".
• ⭐ Ruling 2's literal spelling contradicts its own headline, and I followed the headline. "Accept a callee that is a property access whose base is an identifier" admits promise.then(cb) and rows.map(cb) — promise and rowsare identifiers — which is precisely the blanket widening the same ruling forbids two sentences later, and which the issue body names as the reason a blanket fix is worse than the defect. The issue body's own precise spelling ("it / test / describe followed by a modifier") is the one that is self-consistent, so the implementation is roster-gated. The 209-vs-22 measurement above is what decided it. Flagging rather than silently choosing a side.
• ✅ All five line anchors handed down (:521, :1695, :2451, :2493, and the 132-case count) were correct. Surface measured as claimed: git grep -ln "callbackSiteName" → exactly one file.
• ✅ The 9 unreachable packages/spec families printed by dispatch-gates are the known [finding] 9 packages/spec gate families declare path populations the tree no longer has — a quiet green on every card, printed but never recorded #12514; not re-filed.
• ⚠️Measured on this body itself, twice. GitHub's sanitizer silently ate every short angle-bracket fragment on creation (a path placeholder and two spellings of the anonymous key literal) — inside backticks, and the eaten 'anonymous' left the text asserting the fallback was an empty string, which is false. The first repair moved the literal into a fenced code block and that was eaten too, so a fence is no protection either. Both are now spelled in prose. ⛔ Worth knowing before quoting any angle-bracketed literal in a GitHub body.

Changeset — rule applied, and which one

No changeset. AGENTS.md line 1108: "Pure bug fixes do not require a changeset." This is a repo-internal CI gate script under scripts/, published by nothing and user-visible in no way; the change alters how a gate names sites in its own registry keys, not what it refuses. skip-changeset applied — verified it is a real mechanism in this repo (.github/workflows/pr-automation.yml:294 reads it live, not from the event payload) rather than assumed.

Verification

typecheck is NOT APPLICABLE to a .mjs gate script — the root tsconfig.json sets no allowJs, so tsc --listFiles returns 0 hits for this file. ⛔ Recorded as not-applicable, never as a green.

All gates below run on the final commit 605beb0b3, exit codes captured before any pipe (cmd > log 2>&1; code=$?), under os-verify-lock (VERDICT command-exit 0 · held the lock 73s · waited 1s):

gate	exit
check:cli-test-child-env (self-test 152 pass + live tree)	0
check:agent-test-spelling	0
check:bash32-floor	0
check:cli-command-ids	0
check:cross-package-test-inputs	0
check:entry-guard	0
check:parse-guard	0
check:pnpm-filter-targets	0
check:nul-bytes	0
check:pm-dispatch-gates (convention-triggered)	0
scripts/check-ci-filter-parity.mjs	0
scripts/check-cross-package-test-inputs.mjs	0
scripts/pm/bare-root-worklist.mjs --self-test (convention-triggered)	0

Families derived with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (10 path-derived + the 2 convention-triggered the order named). Control-character self-scan beyond the gate: grep -naP '[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]' → no hits.

---

Generated by Claude Code