Price exact dirty-workspace restoration for QQ16 recovery

[Clear1nsight]

This tests the user-assigned reversible-computation repair of #3985's owned storage obligation. It supplies literal complete recovery words using eight mutable borrowed query banks plus one borrowed scratch bank, all with arbitrary independent content in F16 and returned exactly. The original 224 current X/Y input fields remain frozen; all 224 outputs still finish in F17. No borrowed array is cleared.

The fixed-query-frame strategy restores its borrowed state but costs 976 H1, 1,952 native wrappers, 1,104 maps and 40,400 scalar allowance in the complete decoder. Query compute/uncompute pays 832 native lines; paired responses pay 32; the unchanged source prefix/suffix pays 112. This removes the 112 extra owned D arrays of #3985 but exceeds #3976's explicitly conditional 553-H1 native-only allowance by 423. It is an obstruction for this literal repair, not a general dirty-ancilla lower bound or global supplier rejection.

One bounded pivot temporarily leaves the 104 owned query outputs aligned until uncomputation. Its complete bill is 352 H1, 704 wrappers, 480 maps and 30,416 scalar allowance, with the same exact borrowed return. It is explicitly not an improved decoder: the original canonical #2707 already costs 336 H1 with no extra owned D arrays. The package stops both routes and preserves that comparator.

PROOF.md derives the ordered dirty-baseline commutator, pays every inverse path and explains why the disjoint query/response supports permit uncomputation. WORDS.json freezes both complete words. RESULT.json includes the full invoices and a concrete dirty counterexample: with Q0=constant1 and all other borrowed scratch/query sources zero, omitting the first response leaves future W0=-1/4 when the correct query is zero.

Verification authenticates canonical and retained bytes before inherited execution, regenerates the canonical word, and checks both proposals on independent original and current routed inputs plus nine independent dirty symbols. All final values, complete phase frames, original input return and auxiliary content/frame return are checked. Both signed native recipes are checked on all 112 ports and 512 local columns; every 976/352 occurrence is bound to a checked recipe. These are recipe-basis checks, not 976 independent new native proofs. Six semantic controls reject missing dirty compensation, query uncompute, scratch subtraction, map payload, response sign and an illegal borrowed CLEAR.

The nine borrowed F16 banks are an unmet actual caller premise. No allocation, physical access/movement, dirty-range/precision guarantee, changed productive-family coverage, actual global consumer or new kappa is established. #3976's cap is reused only as its named fixed-coverage screen. A parallel output-aliasing proposal is not a scientific or executable dependency.

Workflow: original joint-query-controller card preserved, priority revision 6, base 20ecf2891d80b5acf18d980b2164fb44ea7faa90. Main remained unchanged at round entry, the pre-pivot checkpoint, and the shared 08:47:40 UTC publication checkpoint. Targeted dirty/commutator PR search and shared Discussion checks have partial coverage. Canonical claim.

Credit ywz65 #2707; Arham Tahir #2681; Ali Romman and retained QQ16/native/geometry/supplier lineage; #3985 for the rectangular cut/owned-copy result; and #3976 for the conditional incidence/cap. Reversible compute/uncompute and dirty-baseline subtraction are standard techniques; the new result is their literal source-pinned native realization and bounded failure decision. Chris R / Clear1nsight, with substantial OpenAI Codex assistance. Separate AI proof/code and independent literal-invoice review found no blocker; no human or formal-kernel review is claimed.

Local validation on CPython 3.12.3 completed: normal and optimized complete replay; all six semantic failure controls; two boundary-test groups covering source/manifest/data/word/type mutation, overwrite refusal and minimal normal/optimized Git-free read-only exports; full registered replay PASS_REGISTERED_RESEARCH_PACKAGE outcome=negative-result. The boundary suite completed in 199.693 seconds. Final metadata/pins and EOF/trailing-whitespace checks pass. An extra EOF blank line in the adapted interpreter was removed with parsed-AST equality checked before publication; the final export replay used those bytes. VERIFICATION.json and REVIEWS.json record exact scope and excluded obligations.

Publication: GitHub-created head 7c962fdfb7fa66f8f963ff3ece1f2aae8c9058f7 on Clear1nsight:research/qq16-dirty-commutator-20261019. PASS_RESEARCH_COMMIT_BYTES verified all 40 package/input paths and canonical source ancestry. The publication changes only 18 new files under this package; base-to-head whitespace validation passes. The late shared checkpoint at 08:58:49 UTC again found canonical main and priority revision unchanged, with no new competing question in the targeted partial overlap check. Hosted CI passed at the exact publication head; completed evidence follows.

Hosted validation completed successfully for exact head 7c962fdfb7fa66f8f963ff3ece1f2aae8c9058f7: run 39226837705, replay job 117976305274. Completed logs identify synthetic merge 5c54461567ad421eb79681c4a5626ecf2c4b096c of that head into canonical base 20ecf2891d80b5acf18d980b2164fb44ea7faa90, CPython 3.12.15, the complete mathematical verifier PASS, two boundary-test groups passing in 128.566 seconds, and PASS_REGISTERED_RESEARCH_PACKAGE ... outcome=negative-result. The committed-inputs-unchanged step passed. The plan ran 265 onboarding tests (11 platform skips), used the same review base, and selected only this package.

The official committed-package exporter and Git-independent integrity checker both passed for the same head and all 40 package/input paths. The archive and completed CI logs are retained outside the frozen scientific package. The PR remains open and unmerged; no scientific package bytes were changed for this CI update.