Research: audit PR #2979's group-A native paths and their binding to the T74 invoice (no new kappa)

[commdata2338]

Change and mathematical scope

Add research/rank7-group-a-entrance-audit/, a source-pinned finite interface audit of PR #2979's group A: helpers 331, 450, 451, 508 and 509, both literal providers and both orientations. No defect was found in this scope; this is partial evidence, not a new supplier admission or kappa.

The added check binds each complete direct scalar/climb/wrapper/private-lifetime path to the literal source, checks all ten normalized helper/private endpoints, and joins the emitted native widths to all 3,330 group-A placements in the final T74 allocation. Helper 331 has width 7; the four helpers entering at [30] have width 6 despite their rank-7 terminal frame. A time-major recount covers 271,728 roles and 149,184 bins, including empty-bin charges. The complete recursive bill is {1:1645680,2:470196,3:159468,4:71928,5:13716,8:5328,48:28190}, with rank mass 4816512 = 65088*74 and nominal stock 66,813.

The fixed-partner local joint/control differences are -H1+20H2-14H3+2H4-2H5 and 3H1+20H2-14H3+2H4-2H5. The credited PR #2993 replay reproduces guide margins +42,541 and -58,943. These surrogate margins are not a physical idle tariff. The final bill includes later construction changes and is not an isolated group-A gain or a legal counterfactual allocation for the single control.

Coverage correction: PR #3038 already checks both residual classes inside its V45 cone. Their exact identities are corroboration. None of the five group-A post wrappers is in that earlier cone; RESULT.json records the precise direct-event coverage difference. The claim's initially incomplete existing-check inventory is corrected here before submission. The retained PR #3038 receipts were inspected, not freshly replayed.

Task and overlap

Task card: research/rank7-group-a-entrance-audit/TASK.json (included in this PR).

• Task / track: independent-moment / independent-verification, priority revision 6.
• Scope key: 5e0f415b25b13c102653; concrete refinement: PR Conditional κ ≈ 0.0017290339826465: joint rank-7 end frames, five paired extractions and #2957's rewrites on #2942's words (+5.00e-6 over #2969) #2979 group-A mixed-entrance native boundary and actual owner-invoice join.
• Question: does the native operator actually supplied to each changed helper have the width charged by the completed invoice, with its entrance, chronology and cleanup retained?
• Scope status: refined. Reuse and credit the existing group/control and residual proofs; add the complete direct boundary and native-width-to-owner join.
• Bounded method/domain: two 704-name source words, all 512 address columns, five helpers/private endpoints, 666 calls and 408 helper slots. New source/event and allocation checks; the retained exact kernel, frame ledger and generic covector engine are explicitly identified as reuse.
• Acceptance: exact source binding, positive identities and semantic rejecting controls, with an exact coverage delta. Stop with explicit unresolved physical/admission obligations; no lowering census or dependent construction.
• Budget/stop: one session of about 60 minutes; no expansion beyond this finite audit.
• Handoff: PR Independently audit #2979's joint rank-seven frame paths #2993's native/physical obligation, after Handoffs and requests for independent checks #1166 and the canonical #2612 claim.

Base: 463113d675af7410ac91d53a4fed51ec6494eb9a; head: 63bb8c58175ef6d97c4dd6407f0802d5f6ed4e54. Canonical main was fetched again at 2026-10-17 17:55:54 UTC and remained identical (still so at 18:43:53 UTC); priority 6, the selected record and scientific pins are unchanged. The final offline overlap search covered 253 open PRs including drafts, fetched 17:49:11 UTC, and 2,017 Discussion records, fetched 17:52:04 UTC, with complete pagination; Discussion bodies were read from the 17:39:04 UTC snapshot. A later snapshot (274 open PRs, 18:32:44 UTC) adds #3208, which proves private zero return on PR #3044's separate R2 words, and #3228, which checks stage-one frame-group additivity and explicitly excludes this native audit; both are complementary. Snapshots are non-atomic and cannot exclude unpublished work.

Exact dependency heads:

Source	Head	Use
PR #2979	be191346391661f4d35fc0d1508d0a6029572576	Literal parent/final words, construction changes and completed allocation
PR #2993	34a7003866390618b8e8006dc483e7a37a37903f	Reused frame ledger and fixed-partner group/control builder
PR #3006	1432a43a331ec857f45093a1d6e16acdbe9849e2	Reused generic reverse-covector endpoint engine
PR #3038	c7d0d7be056ba77fda1c168641b2bcffb10cf43d	Exact retained V45 coverage receipts; no fresh full replay

Nearest additional checks are PR #3008 (948a997eba1aaf13d04e2b52edf70404860d19dc, helper249 event path), PR #3014 (16e8db5656b4fea487a2212395f39707bf134f7a, frozen-parent V45 allocation), PR #2987 (02c3047a1e78e6e35e7573c12434ba26f6f9bd96, final placement and packing geometry, with the width registry as an input) and PR #3079 (8aee45e6a918c99cdca98431c3e791d2b3fd4206, the published T74 aggregate arithmetic without owner chronology). The recount here derives the idle bill 17310 H1 + 4956 H3 + 1728 H5 + 28190 H48, which PR #2993 left as a conditional remainder, from the actual bins with the native-joined widths. No predecessor is modified. The current #2612 all-private-free and V45 final-fill-control investigations are complementary: this package restricts native path checking to group A and does not perform their enlarged closure or counterfactual packing experiments. The PR #3202/#3044 cleanup and PR #3204 idle-axis work concern different objects.

Validation

Python 3.12.3, standard library only. From the repository root:

timeout 30s python3 -B scripts/research_package.py --package research/rank7-group-a-entrance-audit --check-only
timeout 300s python3 -B scripts/research_package.py --package research/rank7-group-a-entrance-audit
timeout 180s python3 -O -Werror -B research/rank7-group-a-entrance-audit/verify.py
timeout 60s python3 -B scripts/check_research_commit.py --package research/rank7-group-a-entrance-audit

Expected main marker: PASS_GROUP_A_ENTRANCE_NATIVE_INVOICE providers=2 helpers=5 residual_widths=7,6,6,6,6 roles=271728 no_new_kappa.

Metadata/pins, the registered replay and all seven behavioral groups pass. The registered run took 27.71 s with peak 340,684 KiB; the optimized replay took 12.45 s. Committed-byte preflight passes for all 33 package/input files, and git diff --check passes. The selected baseline binding check passes; it is not a full selected-result replay.

Controls reject a same-rank wrong entrance, cold-wrapper substitution, global-phase corruption, omitted entrance/wrapper, premature private free, scalar deletion/sign/clock changes, missing cleanup, duplicate/missing roles, a bad capacity/family map, and misuse of terminal dimension as physical width. Source corruption, changed event digests and duplicate JSON keys also reject. All 512 native address columns and global phase are exact. The short support-count argument proves width minimality only within the monomial/Fourier/monomial format.

Not checked: the complete peer-native/warp dependency closure; the other 86 private endpoints; physical stock adequacy, movement, prefix/precision/tape costs; legal whole-supplier parent/joint/single-control allocations; the complex root/packing cap, bit cap, outer coupling/release and inherited all-size/compiler/analytic hypotheses. Nominal stock reconciliation is not physical admission. No Lean, Docker, GPU, large search or full cold supplier replay was run.

Credit

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

Marius-Paul Dumitrean supplied PR #2979 and PR #2942; Dugongue supplied the PR #2969 comparator and PR #2839's owner engine. Nick Petalas supplied the PR #2993 frame/control checker and the PR #3006 covector engine, and PR #3008/#3014 are method predecessors. ToikY (PR #3038) already checks the residual classes; its stronger coverage is explicitly credited. Ross Richardson (PR #2987) supplies complementary packing evidence, and Gonzalo Mier (PR #3079) the aggregate T74 replay. Retain Jeremy Kothe, Mason Hale, Jeremy Diviney, icekylinx, Sagar Thakrar, DreamingOfClouds, chafreaky (PR #458), bbMisk (PR #1611), ToikY (PR #1710), Tekalign, Yiyang Lu, Cinna Davis and the source geometry/owner/native lineage. Existing licenses and frozen source bytes remain unchanged. Parallel #2612 private-free and V45 control work is acknowledged without claiming its results.

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

Start from this group's 55/4/4/5/5 private-touch inventories and bind the complete peer-native and warp dependency closure before interpreting the direct identities as a complete physical caller. Coordinate with the active all-private-free audit first. Then independently review the retained physical/prefix inputs and both supplier caps. Legal comparable whole-supplier control allocations remain a separate obligation; another surrogate evaluation does not supply them. Structured next questions and blockers are retained in TASK.json.

Workflow feedback (optional)

The named next decision and handoff-first workflow focused the session. The catalogue's empty existing_checks required a wider inventory of main and the Discussions, which found PR #3038 only during final coverage review. Populating that field with exact package paths, and with the post-wrapper, endpoint and invoice obligations that remain, would reduce accidental overlap and incomplete novelty claims.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X86sCSXdoDd61SDFfC1PKG