research: price P104 encoded-source return at a retained inverse-shear caller

[Clear1nsight]

Change and mathematical scope

PR #3882 returns encoded residuals in changed physical locations; PR #3894 returns the original sources. This package prices the return needed by the retained two-stage complex caller's inverse-shear equation, rather than treating those boundaries as interchangeable.

Results:

• A fixed-location adapter restores all original X source values, retains every corrected sum, and restores all 72 dirty helpers. It uses 1,152 unit additions, 104 halves, 104 doubles and 192 H1 conversions. The 192-rank tariff is sharp under the stated fixed-register nested dimension-gap model.
• Even granting free logical gathering into residual order (q,v_N), substituting those residuals for x in the inverse shear has rational error rank 104.
• Fusing restoration into the correction gives a source-preserving suffix with 1,632 unit additions, 200 halves, 200 doubles and 208 H1. Fusing the bounded inverse reader gives 1,736 additions at the same 208 H1. A fair retimed/lowered PR3894 control uses 5,096 unit additions, 104 halves, 104 doubles and 208 H1 for the same tested endpoint (-y,x+y).
• Both fused routes retain mass 2,576 and the failed local screen 46/45. The inherited source-preserving nine-stage complete-coverage lower bill is 62,280 > 60,736, or 7785/7592 > 1. No supplier expansion is justified.

The actual retained caller is notes/copied-centers-complex.tex, specifically its inverse second-shear equation and later paid endpoint correction. The package supplies an explicit local common-chart reader with complete frame charges. It does not instantiate that caller's orthogonal-axis geometry or a physical global P104 owner embedding. Those remain open. The older 1,352-add chronological snapshot word remains credited; this is not a record for all source-preserving corrections.

Only research/p104-role-exchange-caller-audit/ is added. No new kappa or selected-frontier change.

Task and overlap

Explicit user assignment, p104-side-correction, changed scope, priority revision 6. Canonical base 20ecf28 was refreshed during the study and again before publication.

Exact scientific sources:

• PR3882: 9d56eee. Its later workflow-only publication child d765863 was inspected separately; frozen scientific inputs remain at 9d56eee.
• PR3894: 13b397e.
• Original channel, fixture and retained caller: canonical base above.

The exact original catalogue card, changed scope and stopping decision are in TASK.json. Entry overlap used 633 open PR records with reported complete pagination, the repository helper against that snapshot, and bounded required Discussion-page reads. Publication retry refreshed all 641 open PR bodies including drafts across seven pages, reran the repository overlap helper, and read bounded live Discussion pages #1163, #1166, #2750 and #2619. A newest-20 delta additionally inspected #3939, a distinct pooled R438 proposal. PR3882 and PR3894 remain at their recorded heads; nearby P104 consumer/flag/lifetime work has distinct contracts. There is no complete Discussion cache, exhaustive historical diff survey or novelty guarantee.

Validation

CPython 3.12.3, Linux, standard library only:

• Primary normal and optimized source-bound replay: PASS.
• Original 4,058-event scalar/frame word on all 280 independent input columns: PASS.
• Eight full caller-cut/return/baseline schedules, exact coefficients and frames: PASS.
• Separate dense quarter-integer reconstruction: 627,200 coefficient comparisons, PASS.
• Nine optimized behavioral groups, including actual Git-free read-only replays and altered/missing-source, typed-receipt and program corruption CLI failures: PASS.
• Export metadata structure, original-card equality and exact canonical source SHA256/Git-blob checks: PASS.

From a complete checkout:

python3.12 -I -B scripts/research_package.py --package research/p104-role-exchange-caller-audit

Direct commands are in README.md. PROGRAMS.json.gz is checked by its complete decompressed canonical payload, not gzip container bytes.

Publication retry verified GitHub commit 51ba88c has exactly the 19 intended added files, with all Git blob hashes equal to the frozen verified package. The primary source-bound replay was repeated and passed. The original source mirror could not run full Git registration or hosted CI. The subsequent upstream hosted replay now passes, as recorded below. The separate repository publication-byte/source-history scripts remain unrun locally; the exact 19-file Git-blob comparison is the publication-byte evidence stated above. VALIDATION.json retains the original study-time limitation explicitly. The installed ruff process segfaulted; no lint pass is claimed.

New native recipe execution, literal signed-floor/native prefix safety, physical owners/routing/tape/holding, the original channel's internal numerical bound, full two-axis caller adoption, all-size admission, selected cold replay, formal-kernel acceptance and independent human review remain unrun or unproved.

Credit

Credit Clear1nsight/PR3882; 1337isnot1337/PR3894; Rohan Arun/PR1972; Chafik Boukhalfa/PR1991; Adam Gall; Josie Elliston/PR1071 and the E8 lineage; the retained correction-rank/factorization/snapshot package; Aurel Prosz/Paureel, Zhihao Chen, Rohan Arun, Dominik Scholz and icekylinx for the actual caller; Sagar Thakrar/PR1012 and Douglas Colkitt for coverage context; Jose Manuel and Vasek/PR2744 for consumer-budget discipline. PR3609/3620's distinct dual contracts are not transplanted.

Prepared by Clear1nsight with substantial OpenAI Codex assistance. Both implementations share one assisted session and source context; this is not independent human review. Apache-2.0 and exact predecessor notices/credit are retained.

Open next questions

README.md and TASK.json release two bounded questions: supply a genuinely matched physical P104 caller with changed complete coverage, or change the two-axis caller and its endpoint correction jointly. Begin with all-role intertwiners, actual owners/returns and the whole bill. Further native/supplier construction stops if the favorable fully charged budget fails.

Workflow feedback

The original publication attempt was blocked before execution by automatic permission-review timeout and capacity. This retry publishes the byte-identical verified research package through GitHub. The integration can write the fork but returned HTTP 403 for upstream PR creation, so this focused submission uses the browser. The contribution adds only the scoped package; no shared CI repair, selected-result edit or broader research claim is included.

Hosted CI confirmation

Checked 2026-10-19 at 05:38 UTC. Registered research packages run 39205163761 completed successfully on its first attempt for scientific head 51ba88c007cfbcc3d9cf22dd321f5224be3a5bb1, tested through PR merge ref f6fcfee75437a0b7da3c25108996d96fa3fee1b2 against unchanged canonical main 20ecf2891d80b5acf18d980b2164fb44ea7faa90.

• Planning job: PASS; 265 onboarding tests, with 11 platform-specific skips, and the changed-package plan selected this package.
• Registered package replay: PASS on CPython 3.12.15 / Ubuntu 24.04.5. All three declared commands completed: primary source-bound audit; separate quarter-integer reconstruction (coefficients=627200 programs=8 fixed_location_rank=192); and all nine behavioral tests.
• PASS_REGISTERED_RESEARCH_PACKAGE ... outcome=partial and final git diff --exit-code: PASS.

No failing CI job or annotation was found. No corrective source, workflow or scientific change was needed. This hosted confirmation supersedes the earlier pending status; it does not establish native execution, physical supplier admission, selected cold replay, independent human review or formal-kernel acceptance. Scoped claims, frozen inputs, predecessor credit and selected frontier remain unchanged.

[Clear1nsight]

DONE / RELEASED — bounded two-axis caller study, published as PR #3954, exact commit f1e2839. This closes the explicit assignment recorded in this comment; no continuing reservation or native/supplier expansion.

Established: the ordered terminal encoding is L_T=[[D,H T],[0,I]], with literal slot encoding A_T=Pi L_T and correction Pi P_encoded Pi; Pi explicitly swaps the 96 N-port X/Y slots and is not a free physical move, and its paid endpoint correction must be conjugated. Applying the old correction has ordered defect [[-H,0],[T^-1(D-I),H]], and directly in literal slots the first-bank defect is -H T^-1. Both are independently checked. A 280-input literal common-chart encoded inverse reader, full operator equations, four-address physical counterexample, all 512 fiber phase patterns, gauge-specific normalization phases and a completed-uniform-transform positive control are supplied.

The named 9-by-9 product-line scaffold has 10,816 data pairs, 36,608 persistent roles and 208 local calls. Its deliberately favorable unchanged profile/progress screen needs 3,009,344 rank units versus 2,965,248 capacity, excess 44,096; unknown added physical work is not zero. This conditional rejection does not supply a global P104 embedding or exclude different joint encodings/owner/coverage laws. The 96 linear-map rank is not a native-tariff lower bound; the phase census is explicit-gauge/one-column data and the particular sign mismatch cancels for even column count.

Evidence: registered Python 3.12 replay, separate noncommutative/rational/translation/owner-set checker, eight behavioral groups, normal/optimized checks, exact 15-file committed-byte/source-history preflight, official export integrity and all three optimized commands in a read-only Git-free export pass. All 11 submitted files match the verified bytes; all 16 exported hashes/modes remain unchanged. Upstream CI is pending separately.

Next released question: provide a changed all-role entrance/exit intertwiner with a favorable complete owner/progress/profile ledger before native work. Completed common transforms commute with scalar bank encodings, so a jointly matched encoded continuation remains open. The old h=28 producer embedding, opposite-orientation native execution, routing/precision/tape, selected cold replay, all-size supplier and independent human/formal review remain unclaimed.

Canonical main was refreshed unchanged at 20ecf28, priority 6. Entry: 650 open PR bodies/seven pages/helper; prepublication latest-100 delta and bounded live Discussion reads. Credit #3882/#3940/#3894 and parallel #3943, Rohan Arun, Chafik Boukhalfa, Adam Gall, Josie Elliston, Aurel Prosz/Paureel, Zhihao Chen, Dominik Scholz, icekylinx, Sagar Thakrar, Douglas Colkitt, Jose Manuel, Vasek and retained lineage. Full pins, assumptions and credit are in the package. Clear1nsight with substantial OpenAI Codex assistance; Apache-2.0. Selected frontier unchanged.