diff --git a/src/v2/workflow/floor_pure_producer_share.dag b/src/v2/workflow/floor_pure_producer_share.dag index 74c4344627b..4c7aa03529e 100644 --- a/src/v2/workflow/floor_pure_producer_share.dag +++ b/src/v2/workflow/floor_pure_producer_share.dag @@ -217,6 +217,79 @@ data floor_cross_claim_pure_producers_warm: List = [ "v2.extdeps.languages.bash_command_fold.bash_fold_lex" ] +// grammar_relation_row_for_emitted HAS NOW BEEN MEASURED ON THE THIRD CONJUNCT, AND IT PASSES. +// It was enrolled on recompute AND sharing before the serve-below-recompute criterion above +// existed, so it stood on two of the three tests this file now requires, and gunbc#10141's +// carrier-overlap wall refused it on a NEIGHBOUR's null result -- it serves +// v2.test.manual.rust_add_emit_translate, a carrier of the refused rust_target_model_core_edges, +// whose refusal was NoMeasuredEffectOverItsConsumers. A null about a different producer is +// evidence neither of a cost here nor of its absence, so the state was UNMEASURED, not fine. +// +// TWO PRESENT-VS-ABSENT PAIRS, dispatched on branches rather than as pull requests because CI +// builds refs/pull/N/MERGE and a PR pair measures two different trees the moment main moves: +// 33701546412/33701544509 and 33705546522/33705544280, each present/absent off one main commit +// differing in this one line. THE ARM IS VERIFIED TO HAVE VARIED rather than assumed: the key +// carries 148 fills over 60 consumer modules in both present arms and is WHOLLY ABSENT from both +// absent ledgers, while the other five keys carry identical fill counts in all four runs. +// +// CONDITIONED ON, and stated because rust_target_model_staging measured clean beside a wider row +// and inherited that row's regression the moment the wider row was withdrawn: both bash warm rows +// plus bash_fold_serialize_node, formal_productions_from_catalog_node and +// bash_fold_relation_row_witness enrolled in every arm, and no rust row enrolled in any. +// +// THE CONTROL IS NOT THE ROSTER-EXTERNAL CORPUS, and that correction is the transferable part. +// Normalising the subject's consumers against rows in no consumer module of any enrolled key +// reports a ~12% win, and it is mostly composition: eval_steps is in required_floor_claim_cost.tsv +// and is host-independent, so the subject's own rows split into those whose steps FELL (the serve +// reached them) and those whose steps are BYTE-IDENTICAL across arms (the serve never reached +// them, same claims, same runs). The second group is a control the change provably cannot have +// touched, and it moves almost as much as the first. Rows elsewhere are an ASSUMPTION about the +// corpus; rows doing byte-identical work are an OBSERVATION about the row. +// +// AGAINST THAT CONTROL THE SERVE BEATS THE RECOMPUTE IN BOTH PAIRS, by roughly a tenth on the +// rows it reaches, with about an eighth of their eval_steps avoided; the two pairs' bootstrap +// intervals overlap and neither admits 1. The estimate is biased TOWARD the null, not toward this +// result: a row the serve reached whose step count happened not to move lands in the control group. +// +// WHAT IS RE-DERIVABLE FROM WHAT IS NAMED HERE, AND WHAT IS NOT. An earlier revision of this note +// said `re-derive with the instrument named`, and no instrument was named: the run ids are the +// measurement's STORAGE and `eval_steps` is its OUTPUT column, neither of which performs the +// control split or the interval. That sentence promised a producer that does not exist, which is +// worse than transcribing a number, because a reader who tries to act on it finds nothing to run. +// Corrected rather than softened (raised as REQUEST_CHANGES by codex, review 59321): +// RE-DERIVABLE from the four run ids plus `required_floor_claim_cost.tsv` alone, by a reader +// with no instrument: join the subject's rows across a pair at identity grain; take the +// consumer set from the present arm's own `[floor-shared-fill] modules=` field; split those +// rows into steps-FELL and steps-BYTE-IDENTICAL; the ratio is the aggregate of the first group +// over the aggregate of the second. That is the decisive comparison and it needs arithmetic, +// not tooling. +// NOT RE-DERIVABLE that way: the bootstrap intervals. Resampling needs an RNG the report path +// does not have, so the intervals are recorded as observed and CANNOT be reproduced from what +// this note names. The claim they support -- that neither interval admits 1 -- therefore rests +// on the original measurement and not on anything a later reader can re-run. +// The row's admission stands on the ratio, which is re-derivable. The intervals corroborate it +// and are not load-bearing for it. +// +// SO THE ROW STAYS, on all three conjuncts rather than two. The refusal that opened this question +// is already gone, and by the same reasoning rather than by exemption: gunbc#10141's overlap join +// is narrowed to refused rows whose verdict records a MEASURED cost, so +// `refused_row_carriers_transfer` answers false for NoMeasuredEffectOverItsConsumers and the wall +// never reaches core_edges' carriers. THAT IS OBSERVED AND NOT ONLY DESIGNED: floor run +// 33703215821 is FloorClean with this key enrolled and core_edges in the refused roster -- same +// roster, same carriers, no refusal. An overlap with a producer whose own measurement found no +// effect either way was never evidence about this one. +// +// ONE OBSERVATION LEFT UNEXPLAINED ON PURPOSE, because an unsupported cause in a carrier is read +// as a lead and spends the next reader's time before it spends their doubt: 12 claims carry +// EXACTLY 1044 more eval_steps with this row enrolled -- a constant, not a distribution, across +// claims of very different sizes. It was pre-registered from pair 1 and reproduced 12 of 12 at +// exactly 1044 in pair 2, so it is structural and deterministic rather than a run artifact. It is +// not roster resolution: install_pure_producer_share decodes both lists once per prepared subject. +// It is not admission diverting calls from the cheaper per-frame memo: eval_pure_named_call is +// reached only when pure_call_memo_key returns None, so the tiers are disjoint by key +// availability. Cause unknown; all 12 route through v2.test.manual.rust_add_emit_translate's +// fixture, which is where to look and is not a mechanism. +// // The claim-forced rows carry claim-independent ARGUMENTS in practice (a production // catalog node, a production identity) but are not nullary, so they fill on first touch // and later claims hit on the content-hashed, fully-verified argument row. Measured owners