Scoped negative result: coupled partial read exchanges on P6k4

[at1as]

PR #699's read-exchange notes leave simultaneous partial exchanges across different entering-copy pairs unexplored. This scoped negative result / research note checks a precisely defined part of that space: all two-pair partial choices on its pinned P6k4 input, plus a bounded three-pair follow-up. No κ improvement is claimed.

All 175,456 two-pair group/mask cases and 27,344 cases across 16 connected triples in the direct-target overlap graph violate a target hyperplane under the inherited frame-floor policy. These are parameterized cases, potentially including duplicate completed rewrites. The triple graph is a search restriction, not a proof of circuit independence.

The primary reduced-basis implementation and a separate explicit-vector-set implementation independently reconstruct eligibility and enumerate every case. Their ordered, per-candidate first-refusal digests agree exactly; the second also records explicit nonorthogonal-vector witnesses. Both reconstruct the known full SW4 word byte for byte. The independent checker verifies its exact rational B map on all 728 input columns and the complete frame profile and helper boundaries.

Reproduce offline

Python 3.11+, standard library only:

python3 -B -I research/complex-read-exchange/verify.py

Expected final line:

PASS: 202800 coupled partial choices rejected by both engines; exact 728-column SW4 control; no new kappa

The scoped package includes the pinned reconstructed input, unchanged inherited SW4 recipe/notices, deterministic evidence, both implementations, 13 mathematical failure tests, source/overlap records, and reconstruction/validation receipts.

Validation and limits

• Fresh Conditional κ = 0.001660666354374641: four entrance promotions, six dim-8 ends and a width-neutral B read exchange on the PR659 word #699/Conditional κ = 0.001636241501576388: h4a plus PR627's two factored jp8 cold ends #635 construction through P6k4 and SW4 passed content pins, native transport, ledgers, local replay, wrappers, inherited mutation controls, and the independent forward/dual 728-column audit.
• Both complete new domains and independent refusal replays passed on Python 3.11.13. The complete public command, including all 13 tests, also passed in isolated optimized Python 3.13.4.
• Scalar corruption, frame corruption, missing compensation, altered helper boundaries, conflicting edits, non-dyadic coefficients, and unfinished/truncated evidence fail closed.
• The selected-record/source-inventory check passes. No supplier, selected certificate, root Makefile or upstream/ file changes. Full repository, phase, bit-supplier and global owner/assembly replays were not repeated for this standalone negative result. Remote CI remains separate.

The exclusion fixes the input, eligible-copy/insertion rules, untouched full SW4 moves, helper identities/boundaries, and frame-floor policy. It does not cover arbitrary triples or larger groups, alternative insertion positions, all-full combinations, relaxed floors, moved boundaries, or different decoders. No candidate passed frames, so no new κ admission was attempted. This is not a proof-assistant formalization or full multiplication-theorem audit.

Base: 66849e1475c930fd7732b1a0067ffee7e1e95f60; #699 source: 03e8d7ca494a083e5053cd6b1228f66b98414479; #635 dependency: a6895a69f8ad98ab45777b03e9e6f48888c39f6f. The final overlap refresh covered 586 open PRs through #767. Concurrent #741 improves full SW4 through eight frame coalescences and is distinct. #699's later f042731c change touches README/MANIFEST only; the construction inputs are unchanged.

Credit to Jeremy Kothe and the BRoute read-exchange construction, Chafik Boukhalfa's owner map and concurrent #741, icekylinx's factory, #589's bit supplier and the retained upstream lineage. Prepared for Jason Willems (@at1as) with substantial OpenAI Codex assistance and separate AI reviews. Apache-2.0; no independent human-review or priority claim. Deliberately selected complex track, OS diversification seed 21; local claim because no pinned sprint issue exists. This is separate from completed PR #666.

[Mister-Meeseeks]

100-PR batch review at d4c60b58d48310ffd5b65bc99ff5b40ed2e59b14.

This remains open after scoped triage. Next review step: Complete 202,800-case fixed-domain replay exceeded the 180-second batch budget. Continue in a bounded dedicated job with exact head/input coverage; no exclusion accepted from a timeout.

This is a specific continuation point, not a rejection or a request to redo already recorded checks. Completed local evidence is retained separately from outstanding whole-head obligations.

Batch review, attribution and evidence. Thank you for the contribution.