PROBE ONLY — DO NOT MERGE: uncensored cost measurement for the 102 budget-refused rows

[gunbai-bot]

MEASUREMENT INSTRUMENT. This branch exists to take one number and is deleted afterwards. 30000ms is not a proposed budget and this must never merge.

Why the data on main cannot answer the question

After #8447 restored the roster, main reports:

passed=8951 known_red_held=367 failed=5
stale_quarantine=0 budget_refused=97     -> exit 1

All 97 budget-refused rows are Interrupted, against one budget (1552ms):

elapsed  min=1553  p50=1554  p90=1692  max=20215

Every one of those elapsed figures is a censored lower bound, not a cost. A row killed at 1554ms has true cost >= 1554ms and may need far more. The floor's own diagnostic offers exactly two dispositions — reduce the row's cost, or move it to a lane that declares its own ceiling — and neither can be chosen from censored data.

The tempting inference is that raising the budget to ~1600ms clears 85 of 97, because 85 rows report elapsed under 1600. That inference is invalid: those rows were killed, not finished, so a higher budget lets them run longer and be killed again one quantum later. Acting on it would look like progress and change nothing.

A finer poll stride does not help either — it tightens the overshoot and reproduces the censoring at higher resolution. Only completion yields an exact cost. (Prescription from tidy-pike-117, adopted.)

What this measures

Raising the per-claim budget to 30000ms lets these rows run to completion, converting censored bounds into measurements. Rows needing more are still censored, but at >=30s rather than >=1.55s — enough to separate the two dispositions.

CORRECTION (my own claim, refuted by my own measurement). An earlier version of this line said 30000 "bounds the run." It does not, and I should have caught it: the deadline is cooperative and fires only at 4096-step poll boundaries, so low-step or host-builtin-heavy work sails past it. I measured the existing 1552ms budget overshooting by up to 19,192ms — inversely proportional to steps/ms across four orders of magnitude. The five near-stepless rows (45,552ms in 7 steps) are the controls proving it.

The honest statement: 30 seconds bounds the intended observation window at pollable execution sites. The hard run bound is the 180-minute workflow timeout. Any post-threshold completion or overshoot is itself part of the receipt, not a failure of it. (Correction owed to the side-chat reviewer.)

The receipt must carry verdict AND cost together

These are enrolled expected-red rows, so completion does not imply passing. Three outcomes are all real and must not be collapsed into an elapsed-only ranking:

• completes PASS -> exact cost, and a stale enrollment (stale-quarantine)
• completes FAIL -> exact cost, and the expected red it was enrolled for is finally demonstrated (held)
• still interrupted -> censored still, but at >=30s instead of >=1.55s

Ranking by elapsed alone would silently merge the first two, which differ in what they oblige.

What it will decide

The exact costs partition the 102 rows (97 refused + the 5 PassedOverBudget failures, whose costs are already exact at 2.6s / 22s / 29s / 45s / 75s):

• near 1552ms -> the budget is miscalibrated, not the rows. 1552 is measured-max + one scheduler quantum from a 1527ms observation on 2026-08-18; a snapshot pin re-derived from a live tree is exactly the oracle DESIGN §5 rejects.
• genuinely expensive -> a real cost defect: reduce it, or declare a ceiling.

Census first, partition by mechanism, fix by class — the method Wave 1 (#8396) used to split 339 name-resolution rows into two causes and clear 164 of them.

Disposition

Branch-local experiment. No dissolution obligation is created because nothing here merges: the branch is deleted once the numbers are recorded, per DESIGN — the canary may run and produce measurements, then be deleted, leaving only the facts.

— sent from sleek-moth-351

[gunbai-bot]

MEASUREMENT TAKEN, AND ITS CONCLUSION IS THAT THE BUDGET SHOULD NOT BE RAISED. Closing. Operator ruling: a witness taking 30 seconds "is basically a failing test" — the tests are valid and passing, and their responsibility needs splitting rather than their ceiling raising.

Run 32175959536 at 30000ms against main at 1552ms:

                    main @1552ms    probe @30000ms
budget_refused           97              10
stale_quarantine          0              76
known_red_held          367             378
failed                    5               3

The 97 undecided rows decompose into three populations, which is what the probe existed to establish and what no censored reading could have produced:

• 76 PASS when allowed to finish. Enrolled as expected-red, no longer red. Stale roster rows the 1552ms ceiling was hiding — they never got far enough to report anything.
• 11 genuinely still red. They demonstrate the failure they are enrolled for once they have time (known_red_held 367 -> 378).
• 10 unfinished even at 30 seconds, pinned at ~30,001ms: the same censoring one level up, true cost still unknown.

The trap this rules out, for the record. Raising the budget alone makes main worse — the 76 move budget_refused -> stale_quarantine, equally red. Retiring the 76 alone is worse still — unenrolled plus interrupted reports as failed, so failed=76. Only a joint change moves the needle, and the operator has now ruled out its budget half. That closes the question rather than leaving it open.

One family root-caused, with evidence. dag_arrow_lambda_witness_test — 13 of the slow rows — parses 3-line sources like fn f(xs: List<Int>) -> Bool { any(xs, p => p == 1) }, but each witness calls dag_grammar(), rebuilding a grammar defined across a 6,230-line module with 134 production constructors. The floor gives each claim a fresh frame by design, so nothing is reused: 13 witnesses, 13 full grammar builds, against a subject worth ~1ms of parsing. That is a cost-shape defect, not a slow test.

It does not generalize to all ten families: effect_reach_test (8 rows) carries zero parse or grammar references and is slow for some other reason. At least two mechanisms are in play and I am not asserting one cause across the rest without measuring it.

Disposition: branch deleted, nothing merged, 30000ms was never a proposed budget. The instrument produced the partition and is now waste — per DESIGN the canary runs, produces measurements, and is deleted, leaving the facts. The facts are above.

— sent from sleek-moth-351