fix(scripts): measure-self-test-floor recognises a ternary process.exit as a failure producer (#15339)

[claude]

Fixes #15339

classifyFloor() in scripts/measure-self-test-floor.mjs awards ROSTER only when BOTH
halves match: a roster comparison and a failure production. The failure half read five
literal spellings (failures.push, a literal process.exit(1), throw new Error,
exitCode = 1, ok(false), none of which covers a ternary exit — an ordinary way to
spell a verdict. So a file could carry a complete, working roster floor whose ablations all
fired and still classify NONE.

Nothing on main is mis-reported today (the one ternary carrier was respelled as an
explicit branch while the card was being filed), so the cost is prospective: the next
Tier C author who floors a file whose verdict is a ternary exit gets a green ablation
sweep and an instrument that still says NONE.

What changed

The criterion is now published in two halves, so a control can test each on its own:

• NAMED — the five spellings above, unchanged.
• TERNARY — the exit code is the verdict: process.exit(COND ? 0 : 1), the same with
  any non-zero failing arm (? 0 : N), and the inverted process.exit(COND ? 1 : 0). A
  failing arm that is a non-zero literal produces a failure exactly as process.exit(1)
  does.

Boundary, deliberate: a bare process.exit(EXPR) is still not a failure producer,
however likely the expression is to be non-zero. That is the accident shape this
instrument exists to tell apart from a handshake — over a self-test that returned early it
is process.exit(undefined), which is exit 0, nothing produced and nothing noticed. Source
text can read a non-zero literal in the failing arm; it cannot read an opaque expression.

Six inline controls come with it — inline, like the rest, so they run on every
invocation and cannot become unrun:

control	fixture	expected
setup	TERNARY_EXIT_GATE carries none of the NAMED spellings	asserted, so the rows below cannot pass for the wrong reason
positive recall	roster floor whose only failure production is ? 0 : 1	ROSTER
arm value	same, failing arm ? 0 : 2	ROSTER
inverted arm	same, ? 1 : 0	ROSTER
boundary	same, dispatch replaced by a bare process.exit(EXPR)	NONE
other half	same, roster renamed to a name the criterion does not know	NONE
reuse pin	the ACCIDENT fixture's own dispatch line still matches the TERNARY half	asserted

The last one keeps the criterion pinned to a spelling measured in this tree rather than an
invented one: the ACCIDENT fixture dispatches with exactly this ternary. It could not serve
as the positive fixture itself — it carries failures.push, so a ROSTER verdict on it would
not isolate the ternary — hence the dedicated fixture, written in the existing convention.

Ablation — the recogniser, measured

scripts/check-regen-pending.mjs as it stood at 7706a50bd (roster floor present, verdict
spelled process.exit(failures === 0 ? 0 : 1)) checked out into the worktree, both
recognisers read against the same bytes:

recogniser	that file	census tallies over the same 170 rows
base (5b2ad1b41)	NONE	162 ROSTER / 7 NONE / 1 COUNT
this PR	ROSTER	163 ROSTER / 6 NONE / 1 COUNT

Mutation proven on disk before reading (git hash-object moved off the HEAD blob; one
ternary-exit line, zero explicit-branch process.exit(1) lines). Restore proven after:
on-disk hash 9f0342101f5f861eab335d1a304cdca7fdf9ca31 equals HEAD:scripts/check-regen-pending.mjs,
git diff HEAD empty, git status --porcelain empty. The restore leg ran from a
trap ... EXIT INT TERM with absolute paths.

Verification (exit codes captured before any pipe)

Against 9ccd3cdc4:

• node scripts/measure-self-test-floor.mjs — exit 0, inline controls pass:
  170 file(s) under scripts/ dispatch on --self-test / 6 of 170. Floored: 163 roster, 1 count-candidate(s) to read.
  Identical before and after the change.
• node scripts/measure-self-test-floor.mjs --json — byte-identical before/after
  (diff exit 0); tallies 163 ROSTER / 6 NONE / 1 COUNT over 170 rows, as asserted. No key
  renamed; the change is additive.
• Probe path untouched: the diff's three hunks all sit in Instrument 1 and the controls.
  probeEarlyReturn run from the base and the patched module over the three handshake
  spellings in this tree (scripts/invoked-as.mjs — a reached-verdict flag;
  scripts/js-comment-mask.mjs — !== SELF_TEST_VERDICT; scripts/check-nul-bytes.mjs —
  the same inside an else if dispatch) returns identical rows, HELD in all six runs.
  A full --probe sweep is CI's; this is the declared narrowing.
• Derived gate family (node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack
  — 20 commands, every one run) plus the always-runs tail (8) plus whole-repo pnpm lint:
  29 of 29 exit 0 at 9d98ef60a, the final commit; the wrapper's own line reads
  VERDICT command-exit 0 · held the lock 643s. Sample verdicts:
  • ✓ dispatch-gates self-test: 1392 cases pass.
  • scope: 233 file(s) under scripts/, 182 carrying --self-test in code (comments masked); 168 of those are run by 30 workflow(s); 164 have their self-test run through the flag, 4 through a recorded route. (check-self-test-wired.mjs)
  • ✓ check:declared-population-live — 201 of 252 famil(ies) declare a path population, and every one of them reaches this tree's 7474 tracked file(s).
  • check-nul-bytes: OK (scanned 7467 text file(s) ... no raw ASCII control bytes).
  • OK check-ratchet-remedy-authority: 202 scripts swept ...
  • pnpm lint (eslint . --no-inline-config) printed no findings, exit 0.

No changeset: this touches scripts/** only and publishes nothing from any package, so the
PR carries skip-changeset.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_012zGPuVVX3deAx9LdjK8jCk

---

Generated by Claude Code