Repository navigation
research: consume P104 residuals in actual cleanup and price source return - #3943
Conversation
|
DONE / RELEASED — the matched return-contract audit is published as PR #3956, exact head f73eef8. Both adapters have identical formal encoded entrance, fixed physical return, charts, unit arithmetic and all 72 arbitrary dirty helpers. PR3943's existing 960-ADD construction saves exactly 192 ADDs against PR3940's 1152-ADD closed adapter; both retain 104 halves, 104 restoring doubles and 192 transports. Two versus four direct updates per nonpivot explains the difference. Recursive rank mass and failed supplier allowance are unchanged; PR3940's fused caller is a different comparison. Evidence: five pinned exports; independently written arithmetic/inverse oracle on all 280 encoded basis inputs; five separate rational block identities; nine semantic controls; normal/-O and read-only Git-free checks; registered Python 3.12 and publication-byte/source-history passes. Actual predecessor schedule builders are reused explicitly. No human/kernel review or native/supplier admission is claimed; hosted CI is separate. Reproduce: python3.12 -B scripts/research_package.py --package research/p104-return-contract-match (about one second locally). Main20ecf289 / priority6 refreshed; 657 open PR bodies including drafts checked before submission, with a newest-body delta through PR3955; wider Discussion coverage remains partial. New PR3954's distinct two-axis scaffold is credited, not replayed. No continuing research reservation. A changed all-role encoding/consumer with a favorable complete owner/progress/profile invoice remains the consequential task; physical/native, precision, routing/tape, opposite-orientation, all-size and selected-cold obligations remain open. Credit Chris R / Clear1nsight PR3882/3940/3943/3954 and the retained P104 channel, terminal-floor, E8, coverage, caller/native lineage. This contribution is the comparative audit, not ownership of the 960-ADD construction. ToikY with substantial OpenAI Codex assistance, including separate AI review/tests; Apache-2.0. Maintainer edits disabled. No merge or frontier change. |
3 participants
Change and mathematical scope
Publish fork Clear1nsight#11 as three science commits at 2cba526: 11 added files only under
research/p104-residual-consumer/. The duplicated P112 workflow repair is excluded and remains separately submitted in #3885. Frozen predecessors and the selected frontier are unchanged.This package makes #3882's residuals an actual consumer of the native final-query cleanup. At the prefix cut before events 3642..4057, X=x, Y=y+Ax and Q=h_Q-Dx; other dirty helpers are restored. With J=[0,1,3,6,22,31,93,103], B=A[J,N] and q=x_J+B x_N, the actual suffix factors as D=T[I B], hence Dx=Tq. Move the unchanged #3882 tail to that cut and replace 416 weighted cleanup ADDs by 32 signed-double updates Q+=Tq. Strict read-only-donor lowering changes 832 unit Gaussian array ADDs to 64, saving 768 array ADDs (786,432 real-component ADDs at 512 complex words). All 280 output rows, prescribed frames and 72 dirty-helper restorations are retained. This is a scoped scalar cleanup saving; no runtime, scalar-optimality, supplier or kappa improvement is claimed.
The separately priced legacy source-return adapter costs 192 formal rank-one transports, 960 unit array ADDs, 104 halves and 104 restoring doubles. Its 192 floor is sharp only under the declared equal-chart, nested-transport, fixed-occupancy contract. Encoded native profile (1282,58,86,104,48,24,8,8) has mass 2576; adding source return yields (1474,58,86,104,48,24,8,8), mass 2768.
For the explicitly hypothetical equal-profile paired chain s=2p+1, m=8s+1, W=104(s-1), granting every stage the same profile, 2312 net endpoint progress and zero other costs, the proof gives M(a)>=1+(160s+104)/(m(W+b))>1 and M_return(a)>=1+(352s+104)/(m(W+b))>1 for p>=1, finite b>=0 and 0<=a<1. The nine-stage paid masses are 62280 and 64008 versus 60736. This rejects that named supplier model; it does not establish physical embedding or exclude arbitrary profiles/caller contracts.
Task and overlap
Base 20ecf28, priority revision 6. TASK.json records task p104-side-correction, the explicitly assigned refined consumer/return question, scope and blockers. Exact #3882 scientific source: 9d56eee (full binding in SOURCES.json); source-retention commit dfc903e precedes the consumer and canonical-binary-subspace correction. Original overlap-helper execution was offline; bounded live PR/visible Discussion checks are recorded, not an exhaustive novelty search. Nearby #3894, #3907 and #3915 address distinct caller/lifetime/flag questions.
Comparison with #3940
PR #3940, scientific head 51ba88c, is directly overlapping parallel work and receives credit.
Both analyze a 192-transport return floor and failed paid budgets; that is shared territory, not an independent novelty claim here. The 960-ADD adapter here and #3940's 1152-ADD adapter are different presented schedules. They are not a demonstrated cost advantage or contradiction: endpoints, frames, allowed operations and accounting must first be matched. #3940 also gives a fused 208-H1/mass-2576 caller; this package's unfused mass-2768 return invoice does not invalidate that fusion. The separate contribution offered here is the source-bound internal cleanup consumer and its stated proof domain. No common global optimality claim is made.
Validation
Reproduce at the scientific head:
Source-bound hosted replay passed at exact head 2cba526 on Ubuntu 24.04.5/Python 3.12.15. It checks metadata/pins, source history, normal/-O primary replay, a separate integer-matrix implementation importing no predecessor/primary executable, 16 failure-control groups, and unchanged source/package hashes. Both implementations replay the actual prefix, complete endpoint and return; the primary also checks actual frames, binary nesting and 96 exports. Result digest: cdf6210cb348566f7647722f05c695d268b220909fb12997d12255817985f43f.
Controls reject omitted/early/duplicate cleanup, sign/donor/output errors, missing doubles/transports, wrong scope/coverage and CLI/typed-result/geometry/source corruption, including under -O. An initial hosted basis-spelling failure is disclosed; the final commit compares canonical binary subspaces in both interpreters, without changing inputs, circuit or result. The local Git-free algebra check is separate from source-bound validation. The exact-head upstream result is now recorded below, separately from these retained receipts.
Fresh native/global-phase and signed-floor admission, physical producer domain, width/routing/control/tape, opposite orientation, global owners, all-size compiler, selected cold replay and full supplier admission remain unproved/unrun. No human or formal-kernel acceptance is claimed.
Passing upstream CI
Upstream run 39206058211 passed at exact scientific head 2cba526: planning, normal and optimized source-bound checks, the separate integer-matrix implementation, all 16 control groups, and committed-input preservation. The research replay job completed in 1m07s; package execution took 10 seconds. The source-bound result digest matched cdf6210cb348566f7647722f05c695d268b220909fb12997d12255817985f43f. No repair commit was needed. This confirms registered execution, with independent review and remaining admission obligations still open.
Credit
Retain RohanArun1972, Chafik1991, AdamGall1940, #3882 and terminal-floor/occupancy lineage; Dominik Scholz's larger-gain model; Sagar Thakrar/1012, Douglas and geometry/native/compiler predecessors; #3609, #3620, #3782 and #3810 retain their distinct scope. Credit #3940 and its retained caller/#3894 lineage for parallel return/caller analysis. All retained licenses/notices remain. Chris R / Clear1nsight prepared this package with substantial OpenAI Codex assistance under Apache-2.0.
Open next questions
See README.md and TASK.json. First match this return adapter with #3940 at identical endpoints, frames and operation accounting before drawing any numerical comparison. For progress, identify a concrete caller that consumes the encoded endpoint without unaffordable return, then replay its complete dirty-state restoration and paid ledger. Physical opposite-orientation embedding and a changed supplier profile remain blockers; expanding the rejected equal-profile model alone does not resolve them.
Community uptake and coordination
The separate P104 community handoff includes source bindings, evidence, scope limits, reproduction/runtime, credit and concrete next tasks. The first bounded task is to match the #3940 return/caller contracts before comparing costs; another is to specify and fully price an actual consumer that retains the encoded endpoint. Independent review of the cleanup factorization and named budget proof is also welcome. These questions are available for community uptake without a new reservation from this update. Reply there with a bounded scope and investigation link, check live overlap/claims first, and report findings here.