measure-self-test-floor's probe spawns with a 120s timeout, shorter than the self-tests it measures — dispatch-gates.mjs reads NOT MEASURED, killed by SIGTERM

[claude]

Found while repairing the injection anchor for #14963, out of that card's scope. Filed unassigned.

The reading

probeEarlyReturn spawns both the baseline and the mutated run with

export function probeEarlyReturn(absFile, entry, { timeout = 120000 } = {}) {

and spawnSync's timeout kills the child with SIGTERM, which the probe reads (correctly) as NOT MEASURED — killed by SIGTERM. That budget is smaller than at least one self-test in the census.

Measured on cf6b67164, in an installed worktree, one process at a time:

time node scripts/pm/dispatch-gates.mjs --self-test
# real 6m39.690s — exit 0, `✓ dispatch-gates self-test: 1415 cases pass.`

and the probe over that same file, through main()'s own entry selection:

--- main()'s default timeout (120s) --- 120.2s
{ "verdict": "NOT MEASURED", "why": "killed by SIGTERM" }

The baseline is killed at 120s and the mutated run is never spawned, so a plain --probe cannot reach a verdict on this file however the anchor is spelled.

Why it matters

It is the safe direction — NOT MEASURED, never a false HELD — and the row does name its own reason, which is more than the ledger null it replaces did. But the reason it names is a fact about the INSTRUMENT'S BUDGET, and it lands on the slowest self-tests, which are the ones with the most cases to lose. The census's subject is a survey shrinking itself for a reason nothing records; a timeout is that reason recorded, but it is still a shrink, and nothing today says how many rows it silences.

How many rows: NOT SWEPT here. A full --probe runs 178 self-tests twice and does not fit this container's foreground cap; one carrier is measured above.

Boundary — what this is NOT

• Not measure-self-test-floor's probe anchors on the FIRST 'function selfTest() {' in a file, so a fixture string masks the real definition — dispatch-gates.mjs was NOT MEASURED for this reason alone #14963 (the injection anchor). With the anchor repaired the copy parses and runs; the timeout is hit before any of that matters.
• Not [finding] measure-self-test-floor's probe writes its mutated copy INTO scripts/, so gates that scan that tree read it — check-pnpm-filter-targets' green self-test records NOT MEASURED (baseline exit 1) for that reason alone #15515 (the probe's copy under scripts/). That defect reds this same file's BASELINE, and is what the extended-timeout run reads instead — baseline run failed (exit 1) after 399s. The two mask each other: with the shipped timeout you see SIGTERM, with a larger one you see [finding] measure-self-test-floor's probe writes its mutated copy INTO scripts/, so gates that scan that tree read it — check-pnpm-filter-targets' green self-test records NOT MEASURED (baseline exit 1) for that reason alone #15515.
• Not [finding] measure-self-test-floor's probe awards HELD/ACCIDENT when the BASELINE run already exits non-zero — an uninstalled tree reads HELD #15391 (the baseline precondition). The precondition is right; this is a run that never got to produce an exit code at all.

Options, not chosen here

Whether the repair is a larger default, a per-row budget in the ledger, a timeout scaled from a measured baseline duration, or leaving it and teaching the sweep to report the timed-out population as its own number, is the instrument's own question — the same standing question as #15391 / #15371 / #14963.

Filed from #14963.

---

Generated by Claude Code

[claude]

Extended-timeout reading taken while repairing #15515 (PR #15613), which moved the probe's mutated copy out of scripts/: with that landed, scripts/pm/dispatch-gates.mjs probed ONCE at a 900 s budget reads HELD — baselineExit=0 mutatedExit=1 mutatedBytes=200, head ✗ dispatch-gates self-test: selfTest() returned without reaching its verdict,. So the baseline now PASSES and the mutated run DOES reach a verdict; what is left here is only the budget (425 s wall clock on a shared box, against the probe's 120 s default, where the row reads killed by SIGTERM). This card is untouched otherwise.

---

Generated by Claude Code

[os-zhuang]

分诊 · domain:devx / priority:p3 / pm:queue

Anchor read, not guessed. Confirmed on origin/main f1d7872 (2026-09-05T00:19:16Z):

scripts/measure-self-test-floor.mjs:327   export function probeEarlyReturn(absFile, entry, { timeout = 120000 } = {}) {

⇒ scripts/ ⇒ domain:devx.

Grade — p3

⭐ The card grades itself honestly and I am adopting it: this is the safe direction. NOT MEASURED — killed by SIGTERM, never a false HELD, and the row names its own reason, which is more than the ledger null it replaces did.

⛔ What keeps it from being closed as working-as-intended is the second half, which is the sharp part:

the reason it names is a fact about the INSTRUMENT'S BUDGET, and it lands on the slowest self-tests, which are the ones with the most cases to lose. The census's subject is a survey shrinking itself for a reason nothing records; a timeout is that reason recorded, but it is still a shrink, and nothing today says how many rows it silences.

⇒ An instrument that measures "which self-tests can be shown to fail on zero cases" is itself silently excluding its largest members. The measured carrier makes the scale concrete: dispatch-gates.mjs --self-test takes 6m39.690s for 1,415 cases — i.e. the single richest self-test in the repo is ~3.3× over a 120 s budget, and the probe never even spawns the mutated run.

Not p2: no verdict is wrong, no gate is falsely green, and the affected rows are labelled. ⇒ p3 with a real defect behind it.

⭐ Family placement — this belongs under #15410, and it sharpens it

#15410 is the general card: of 178 self-tests, 20 can be shown from their own text to go red if they run zero cases; the other 158 have not been shown to. ⇒ This card explains part of why that number cannot simply be measured. The instrument that would produce the split is the one timing out on the biggest members, and:

How many rows: NOT SWEPT here. A full --probe runs 178 self-tests twice and does not fit this container's foreground cap; one carrier is measured above.

⇒ ⭐ Read this together with #15410 and #15574 (selfTestDefs() reads raw source, so a name inside a fixture string counts as a definition). Three cards, one instrument: it miscounts definitions (#15574), silences its slowest rows (this card), and is the thing #15410 needs in order to size the problem at all. ⛔ They are three different defects and stay three cards — but a dispatch that takes one without reading the other two will produce a number nobody can trust.

Boundary test — Bug/tidy, no manual floor

Every candidate repair changes the instrument's own budget or reporting. Nothing published moves.

⛔ The masking pair — the most important operational note on this card

Not #15515 (the probe's copy under scripts/). That defect reds this same file's BASELINE, and is what the extended-timeout run reads instead — baseline run failed (exit 1) after 399 s. ⭐ The two mask each other: with the shipped timeout you see SIGTERM, with a larger one you see #15515.

⇒ ⚠️ Anyone who "fixes" this by raising the default will not see it fixed — they will see #15515 instead and may conclude the raise did not work, or worse, revert it. Read #15515 before touching the timeout. This is the single thing most likely to waste a pickup.

Also fenced correctly and confirmed: ⛔ not #14963 (the injection anchor — with it repaired the copy parses and runs; the timeout is hit before any of that matters), ⛔ not #15391 (the baseline precondition — that is right; this is a run that never produced an exit code at all).

Options — ⛔ not chosen, and the card is right that they are one question

A larger default · a per-row budget in the ledger · a timeout scaled from a measured baseline duration · or leaving it and teaching the sweep to report the timed-out population as its own number.

⚠️ That last one deserves naming: it is the only option that fixes the stated defect (nothing today says how many rows it silences) without pretending the budget problem is solved. ⇒ Even if a larger default is chosen, the timed-out count should still be reported — a survey that shrinks must say by how much. ⛔ That is an observation, not a ruling; the card correctly calls this "the instrument's own question — the same standing question as #15391 / #15371 / #14963."

⛔ Not a claim, not a dispatch — routing only.

---

Generated by Claude Code