Audit P104 private-copy lifetimes and price a copy-free query route

[Clear1nsight]

Result

The eight inherited P104 private copies deliver nonzero arbitrary-dirty query rows to 32 real delivery helpers. This audit binds every creation, read, erase, downstream read and cleanup to the frozen word. Removing the eight intervals breaks all 32 delivery-helper returns and leaves a rank-eight defect in the 104 sum outputs after the actual PR3882 continuation.

A concrete copy-free replacement descends the corresponding Q query, performs the same delivery additions and restores Q. It preserves all 280 logical columns and 72 arbitrary dirty helpers, removes eight zero operations, eight copies, eight private erases and one peak local private array, but adds 56 native rank units: channel mass 2368 to 2424 and unchanged PR3882 combined mass 2576 to 2632. Eight exact reverse rank-seven recipes are included. This is a local workspace/rank tradeoff, with no runtime or recursive saving claimed.

The fixed direct-Q route's descent/return tariff is sharp under the same owner, delivery chart and later consumer. It is not a universal COPY lower bound. Under the unchanged PR3954 scaffold hypotheses the already failing excess grows from 44,096 to 55,744. Further dependent construction stops at that failed gate. No selected-frontier change or new kappa.

Scope and verification

Exactly 11 added files under research/p104-copy-retirement/, final head eeb4ae73ff0d6b049fb72608074be78d623cb51d, tree f9cbed27ea11164321637265d5e870952275ff30; canonical base 20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority revision 6.

• New sparse rational row/frame evaluator checks the complete original and replacement words, exact -q_j defects in every affected D helper, and the unchanged actual PR3882 tail. Only the predecessor's geometry and schedule builders are reused; its arithmetic evaluator is not imported.
• Eight new positive rank-seven return recipes: independent retained native interpreter checks exact physical-event binding, global phases and 8,192 Gaussian columns across both directions, plus a phase mutation.
• Four owner/frame/rank faults and seven real optimized-mode CLI controls reject Boolean/float receipts, false mass, duplicate JSON, changed/missing witness and altered tail source.
• Standard-library Python 3.12 normal/optimized registered replay and exact public-commit byte/source-history preflight pass for all 16 declared files. All five commands also pass from the official read-only Git-free export; all 17 exported hashes and modes stay unchanged. Final-head hosted CI passes, as recorded below.

Reproduce:

python3.12 -B scripts/research_package.py --package research/p104-copy-retirement
python3.12 -B scripts/check_research_commit.py --package research/p104-copy-retirement --check-source-history

The source-complete archive at this head has SHA256 4d208eb9627faa4cafdadcdf5e3e935e82cd434c67b7425c967af41f4b21295f; 16 declared files plus its transport receipt. The unchanged PR3882 schedule export is retained in the package, so cold replay does not depend on an unmerged Git object.

Selection, limitations and credit

Assigned bounded successor to PR3954, with changed scope under p104-side-correction. Canonical three-scope claim. Main was refreshed at entry and before publication; all 662 open PR bodies including drafts were checked at entry, then the latest 100 updated PRs and the repository helper were refreshed. Discussion coverage is bounded and coordinated through that claim. Sibling #3961 encoding and #3962 coverage audits are separate; PR3956's completed adapter comparison is not repeated. The latest canonical refresh remained20ecf289; no new same-scope external PR was observed.

Full inherited native-channel replay, signed floors/precision, all-size implementation, actual global ownership, wrapper execution/memory/routing/control costs, opposite orientation and supplier admission remain unrun or unproved. A changed producer, consumer, encoding or coverage contract can escape the restricted direct-query conclusion, but needs its own complete invoice before expansion. The next useful test is one changed delivery consumer that preserves all actual dirty Q/D contributions and avoids repaying the same descent/return bill.

Credit Rohan Arun/PR1972 for the complete COPY56 word and dirty cleanup; Chafik Boukhalfa/PR1991 for its finite native realization; Adam Gall/PR1940 for 32-chain delivery; Chris R/Clear1nsight for PR3882 and PR3940/3943/3954; ToikY/PR3956 for the matched return audit; and the retained Josie Elliston/E8, Sagar Thakrar, Douglas Colkitt, Dominik Scholz, Clifford compiler and PR1862 interpreter lineage. New audit/rewrite/return recipes prepared for Chris R/Clear1nsight with substantial OpenAI Codex assistance and a separate coordinating-agent dirty-column check. Apache-2.0. No independent human or proof-assistant review is claimed. No merge, closure or selected-frontier change.

Final-head hosted confirmation

Upstream run39214945487 passes on first attempt at exact head eeb4ae7, tested at mergeb21c016c4fca7bec65eb4718a6910096c66b0969 against unchanged main20ecf2891d80b5acf18d980b2164fb44ea7faa90. Planning, all five declared commands and git diff --exit-code pass on Python3.12.15. The logs confirm both normal and optimized full row/tail audits, both8,192-column native return checks with global-phase control, and all seven optimized CLI mutations. The source-export and local publication receipts remain separate retained evidence. This execution result preserves the stated partial scope; physical/global, precision, recursive and human/formal obligations remain open.