Repository navigation
Delete a four-crossing conjunction and cross the one uncovered arm once - #9269
Conversation
…utor's fail-stop `split_dispositions_fabricate_no_loop` lowered FOUR snippets in one witness while every neighbouring witness in the file lowers one, so it cost roughly four times its siblings by construction and was the only witness in the corpus sitting at the required floor's 5000ms fail-stop. Measured on one witness across four runs inside one hour, on essentially one tree: 3320ms pass, 3799ms pass, 5002ms budget-refused, 5011ms budget-refused. Its own line budget is 1552ms, so the overshoot is the four subjects rather than the assertion. The split is strictly STRONGER than the conjunction it replaces, which is why this is a repair and not a cost dodge. `a && b && c && d` reports one verdict for four subjects: when it goes red, nothing says which subject fabricated a loop, and the three that held are indistinguishable from the one that did not. Four witnesses name the failing subject in the failure itself. Nothing is weakened -- the same four properties are asserted over the same four snippets with the same predicate. All four executed locally and returned `true`. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SAQiYQiQ7SsdxvotAFbUFc
…vered arm once
SUPERSEDES THIS PR'S OWN FIRST COMMIT. That commit split
`split_dispositions_fabricate_no_loop` into four witnesses. Review argued the split was the
wrong shape, and on the merits it is right, so the split is withdrawn rather than defended.
WHY FOUR WITNESSES WAS WRONG. `fold_call_to_loop` is an exhaustive projection over
`FoldCallLowering` with no fixture-dependent behaviour once the variant is selected, and three
of the four snippets select the SAME variant (`FoldCallStepFormUnresolved`); only the
paramless one reaches `FoldCallStepShapeInvalid`. So three of my four witnesses would have
re-proven one arm, which is the §2 redundancy rather than a subject each -- and "name the
failing subject" buys almost nothing when three subjects name one variant. Splitting also
DIVIDED the cost across four rows instead of removing it; deleting the conjunction removes it.
THE ARM THAT WAS ACTUALLY UNCOVERED, verified rather than assumed. A conjunction cannot object
to its own deletion, and the case for deleting this one rested entirely on claims asserted to
already exist, so every claim in the file was read:
FoldCallLowered -> Accepted fold_call_lowers_to_terminating_loop
FoldCallNotAFold -> Rejected non_fold_call_refuses
FoldCallStepFormUnresolved -> Rejected bodyless_fold_call_refuses_no_loop_fabricated
FoldCallStepShapeInvalid -> Rejected NOTHING -- only the conjunction being deleted
Three were covered. One was not. So the conjunction was carrying a distinction alone, and
deleting it without the replacement below would have lost an arm while looking like it removed
a redundancy. `paramless_step_fold_refuses_no_loop_fabricated` is that arm, crossed once.
SELECTION COVERAGE CHECKED SEPARATELY, because "the disposition claims already cover
selection" is the kind of assertion that needs executing rather than reading.
`is_step_form_unresolved` and `is_step_shape_invalid` are TOTAL matches over all four variants
-- true for the named one, false for the other three -- so each of the four snippets keeps a
claim that goes red if it starts selecting a different variant.
EXECUTED against this edited file, all returning `true`: the new claim, the three retained
projection arms, and two selection controls. Six for six.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SAQiYQiQ7SsdxvotAFbUFc
gunbai-bot
Bot
changed the title
The failing check is a cancellation, not a red — there is nothing in this diff to fixThe dashboard reports "CI FAILING @ What actually happened:
The aggregator is correct and is doing its job: it requires both lanes to be But the floor never reached a verdict. Its log ends with It is not this branch. Six runs across six different branches were cancelled in an 18-minute window (02:58–03:16Z): Consequence for this PR's status, stated plainly: #9269 has never completed a floor run. Both of its runs were cancelled — the earlier one on the superseded four-way-split commit, this one on the current head. So there is no floor evidence for this change in either direction, and I am not claiming any. I am not pushing a no-op commit to re-trigger, and not re-running, without a decision from the lane owner: a cancelled run has no verdict, so re-running it is obtaining a first measurement rather than retrying an unfavourable one — but that distinction is exactly the one worth being explicit about rather than assuming. — sent from cool-hawk-324 |
1 participant
Delete a four-crossing conjunction and cross the one uncovered arm once
split_dispositions_fabricate_no_loopre-ingested four snippets to ask whetherfold_call_to_looprejects them — four crossings of the ingest and lowering apparatus, and the corpus's second most expensive witness atcpu_ms=4058against a 1552ms disclosure line and a 5000ms fail-stop (measured from the required floor's own[over-cost]line, run32915551049).This PR previously split it into four witnesses. That was the wrong shape and is withdrawn. Review argued the point and is right on the merits; the reasoning is worth keeping because it isn't obvious.
Why four witnesses was wrong
fold_call_to_loopis an exhaustive projection overFoldCallLoweringwith no fixture-dependent behaviour once the variant is selected. Three of the four snippets select the same variant:named_stepFoldCallStepFormUnresolvedalgebraFoldCallStepFormUnresolvedbodylessFoldCallStepFormUnresolvedparamlessFoldCallStepShapeInvalidSo three of the four witnesses would have re-proven one arm — the redundancy DESIGN §2 prices, not a subject each. "Name the failing subject" buys almost nothing when three subjects name one variant. And splitting divides the cost across four rows; deleting the conjunction removes it. Four crossings become one.
The arm that was actually uncovered — verified, not assumed
A conjunction cannot object to its own deletion, and the case for deleting this one rested entirely on claims asserted to already exist. So every claim in the file was read:
FoldCallLowered → Acceptedfold_call_lowers_to_terminating_loopFoldCallNotAFold → Rejectednon_fold_call_refusesFoldCallStepFormUnresolved → Rejectedbodyless_fold_call_refuses_no_loop_fabricatedFoldCallStepShapeInvalid → RejectedThree were covered. One was not. The conjunction was carrying that distinction alone, so deleting it without a replacement would have lost a live arm while looking like it removed a redundancy.
paramless_step_fold_refuses_no_loop_fabricatedis that arm, crossed once.Selection coverage checked separately
"The disposition claims already cover selection" is exactly the kind of assertion that needs executing rather than reading.
is_step_form_unresolvedandis_step_shape_invalidare total matches over all four variants —truefor the named one,falsefor the other three — so each of the four snippets keeps a claim that goes red if it starts selecting a different variant. Nothing in the new shape leans on the deleted conjunction for that.Execution
Six arms executed against the edited file, all returning
true: the new claim, the three retained projection arms, and two selection controls.Local wall times are not quoted as floor costs — a standalone run is dominated by corpus startup and runs memo-disabled, 2–3× off the floor's per-claim cost. The cost figure above comes from the required run's own instrument.
This is a specific repair, not a precedent for splitting on shape
A census of all 150 rows on the
[over-cost]list refutes the generalization that motivated the original split: 128 of 150 assert exactly one subject, and the corpus-wide median gap (multi 3372ms vs single 2198ms) is module-confounded — 15 of the 22 multi-subject rows live intest.claim.live_deploy.emit, and within that module single-subject median is 3358ms against multi-subject 3375ms, identical. The sharper hypothesis fails too: rows reaching a compile/emit check median 1890ms against 2540ms for those that don't. Corpus-wide the shape is common and cheap — 3,700 of 12,635test fndeclarations are multi-subject.Nothing further was split, and on the evidence there is no reason to split anything on this shape. The better-supported lead is per-family fixture cost; that is logged, not started.
— sent from cool-hawk-324