Research: audit all PR #2979 private frees and their expanded native dependencies

[commdata2338]

Change and mathematical scope

Close PR #3038's first open question for the original PR #2979 words: all 91 private buffers are zero at their actual free events in both providers and both orientations, for arbitrary original source, target and dirty-helper inputs. The package is research/rank7-peer-private-free-audit/.

The independently implemented forward value/chart/lifetime closure retains the original helper249/private953 roots and adds every private free. Per provider, coverage grows from 8,902 to 12,693 events forward and from 12,357 to 20,956 in inverse transpose. These are transitive producer sets, not coefficient supports: no algebraic cancellation prunes a dependency, so the expanded cone is a conservative closure, not a minimal set of required identities. All 5,815 expanded native classes per provider are checked on all 512 address columns, including global phase. That includes 609 newly reached forward classes and 1,377 newly reached inverse-transpose classes; previously covered classes are replayed too.

A concrete failure control shows why the extension matters. Changing forward cleanup event 17256 from Z395 -= L5 to Z395 += L5 leaves the original two roots correct but leaves Z395 = -2(x64+x65+x96+x97) at free event 22501. Event 17256 is outside the old cone. The unchanged words pass. An inverse-transpose counterpart, premature frees and new-class phase mutations also reject.

This is a finite audit of the conditional normalized macro/native boundary. It makes no word, allocation, tariff, supplier or selected-exponent change. All 91 private families remain charged, with peak 91. The complete local profile is independently recounted from events; the source-declared 66,813-family stock and complete recursive invoice are reconciled arithmetically. This does not revalidate owner geometry, physical scalar/COPY/CLEAR, storage/precision/tape, both supplier caps, the full outer release or inherited all-size hypotheses. Apart from the retained helper249 endpoint, the 632 original supplier output endpoints are outside this audit.

Task and overlap

• Task: independent-moment, track independent-verification, priority revision 6.
• Scope key: 5e0f415b25b13c102653/PR3038-PR2979-all-private-frees.
• Base: 463113d675af7410ac91d53a4fed51ec6494eb9a.
• Scientific construction: PR #2979, be191346391661f4d35fc0d1508d0a6029572576.
• Immediate predecessor: PR #3038, c7d0d7be056ba77fda1c168641b2bcffb10cf43d.
• Prior frame/invoice records consumed: PR #2993, 34a7003866390618b8e8006dc483e7a37a37903f; PR #3014, 16e8db5656b4fea487a2212395f39707bf134f7a.
• Earlier native/event context: PR #3006, 1432a43a331ec857f45093a1d6e16acdbe9849e2; PR #3008, 948a997eba1aaf13d04e2b52edf70404860d19dc.
• Claim: discussion #2612.

Task card. Scope is refined to one named remaining implementation implication. Existing coverage was inspected on main: PR #2993 checks the groups and losing local controls; PR #3014 prices the frozen-parent owner assignment; PR #2987 checks final packing; PR #3006, #3008 and #3038 progressively cover two native endpoints. PR #3038 explicitly leaves peer-private frees open. The question is whether those actual free obligations hold and what additional native dependencies they require. The bounded domain is two literal providers, two orientations, 91 private allocation/free pairs and their fixed event streams. The method is exact rational/free-word propagation and forward transitive provenance, followed by source-bound native class replay. Success is zero at every actual free plus checked native dependency closure; falsification is a nonzero private, illegal lifetime or source/native mismatch. The milestone is an audit, and the stop rule excludes another lowering or packing search. The full cost budget stays unchanged; physical arithmetic, ownership, precision and supplier/outer admission remain separate. The full pinned catalogue card and structured remaining question are in TASK.json.

The new checker imports neither the primary scalar replay nor the predecessor covector/graph algorithms. PR #3038's source decoder and native selector, and Sagar Thakrar's exact Gaussian-integer Clifford kernel, are explicitly reused; their implementation is not claimed as new. Inputs are source-hash and Git-blob bound. Ten consumed PR #2979, #3038, #2993 and #3014 files were also compared directly with their exact dependency heads.

Nearest concurrent work is complementary. The public claims on PR #2979's mixed-entrance group A native residual width, on the final T74 allocation's V45 terminal-restoration controls, and on PR #3153's T20 coupled-inventory obstructions concern other obligations. The concurrent PR #3208 (24kHandsome1201, using PR #3202's artifacts) proves all 91 private zero returns for Nick Petalas's PR #3044 R2 words. Those are different words: PR #3044 rewrites the restore, and its 5,208-gate cleanup shares only its last 5,166 gates with PR #2979's 5,207-gate cleanup. PR #3208 describes this audit as complementary. This account's PR #2655 pools a different PR #2269 word and is not repeated. The task helper and semantic queries were run against the paginated offline overlap snapshot, and current-main archives and discussion #2612 were inspected. An offline search is not a proof of novelty.

The pre-PR refresh at 2026-10-17 17:51:25 UTC and final fetch at 18:00:26 UTC found canonical main unchanged at the base above, with priority revision 6 and no guidance, selected-record or source delta. The final offline helper queries at 18:01 UTC used all 253 open PR bodies fetched at 17:49:11 UTC and 2,017 Discussion records fetched at 17:52:04 UTC, with complete pagination. A packaging recheck at 18:42 UTC found main still at the base; offline queries against 274 open PRs (18:32:44 UTC) and 2,047 Discussion records (18:34:20 UTC) added only PR #3208 as nearby work on this obligation. Offline warnings and absent arbitrary historical diffs/unpublished work remain limitations.

Validation

Environment: Linux, CPython 3.12.3, numpy==2.3.5; one numeric thread. New proof calculations and the exact kernel use the standard library. Producer scripts run from fresh temporary directories with assertions enabled; the new checker runs with -I -O -B and explicit exceptions.

python3.12 -B scripts/research_baseline.py
python3.12 -I -O -B research/rank7-peer-private-free-audit/test_verify.py
python3.12 scripts/research_package.py --package research/rank7-peer-private-free-audit --check-only
python3.12 scripts/research_package.py --package research/rank7-peer-private-free-audit
python3.12 scripts/check_research_commit.py --package research/rank7-peer-private-free-audit

PASS: source baseline bindings, all eight optimized behavioral tests, package metadata/pins and the complete fresh registered replay. The registered replay ran 2026-10-17 17:49:04–17:57:33 UTC in 508.915 seconds, with both provider summaries reporting private_frees=182 native_classes=5815, followed by PASS_REGISTERED_RESEARCH_PACKAGE. The final committed package/input comparison passed for all 28 files at commit 3c689aac34ff240f82e27dcc7f0247b96b8676af; git diff --check passed. All executable bytes were unchanged after the successful replay; the final README, TASK and check descriptions record its outcome. A packaging rerun of the committed package, in a separate CPython 3.12.3 environment with the same pin, passed the complete registered replay in 584 seconds and left the worktree clean.

The package has eight handwritten behavioral tests. Actual-stream controls flip an uncovered cleanup coefficient, free a private early, and corrupt the phase of a newly reached native child. The 90-versus-91 private-cost control rejects. Both frozen-parent joint/single losing controls remain source-bound context, not a new owner audit. The deterministic receipts retain every private free, complete closure ranges, new native class associations, operator-binding digests, exact counterexample rows and the invoice trace.

The first cold attempt exposed a tuple/list mismatch in the in-memory control receipt after the base mathematical checks completed. The control's stored frame is now normalized to its JSON representation. The original evidence values did not change; the fresh final replay tests the full receipt comparison.

No full selected/supplier/outer replay, Lean, Docker, hosted CI or independent human review is claimed.

Credit

Prepared for commdata2338 (@commdata2338), with substantial OpenAI Codex assistance.

Credit ToikY for PR #3038's source/native machinery and explicit peer-free handoff; Nick Petalas for PR #2993, #3006, #3008 and #3014; Marius-Paul Dumitrean for PR #2979 and #2942; Jeremy Kothe for frame/native/cleanup and rewrites; Mason Hale (PR #2917) and Jeremy Diviney (PR #2908) for the extraction chain; Sagar Thakrar for the exact Gaussian-integer Clifford kernel; icekylinx for the native compiler; Ross Richardson for algebra and packing audits, including PR #2987; and Tekalign, Yiyang Lu, Dugongue, Cinna Davis, DreamingOfClouds and the retained geometry/owner lineage. Parallel work: 24kHandsome1201's PR #3202 and PR #3208 on Nick Petalas's PR #3044 words. Existing frozen sources and notices remain unchanged. New material is Apache-2.0.

AI assistance: submitted by @commdata2338 with substantial AI assistance. OpenAI Codex (GPT-6 Astra) did the derivation, the implementation, the checking and the writing; it was packaged and submitted with Claude Code. No independent human review is claimed.

Open next questions

Join the checked 91 private lifetimes and reached scalar/COPY/CLEAR events to the literal final fill and their physical backing, arithmetic, storage and precision charges. This is the smallest next step toward supplier admission; exact local zero-return and native phases alone do not establish that contract. The complete supplier/outer release and inherited all-size/compiler obligations remain open. See the package README and TASK.json; no further fixed-word expansion is proposed for these private-free roots.

Not checked here: whether the zero returns also hold for arbitrary live arrays at the cleanup entrance, the form PR #3208 proves for PR #3044's cleanup. If they do, this obligation needs only the cleanup's own operations, not the prefix part of the expanded cone.

Workflow feedback (optional)

The card's existing_checks is empty even though main already retains PR #2993, #3006, #3008, #3014 and #3038. Following their explicit remaining obligations avoided repeating a covered rank or root calculation. Add these coverage records and PR #3038's now-checked all-private-free boundary to the card, retaining physical admission as open.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X86sCSXdoDd61SDFfC1PKG