[dillona]

The center-free producer has eight code modes, but these do not provide an eight-call Gaussian repair. This package proves sharp 112-call bounds for the actual one-sided correction and eight-mode extraction/expansion, even with chronological linear feedback, saved responses, fanout, free mode rebasing and cross-block mixing. Homogeneous direct sums require 112m calls. The conjugated adapter J and J-I also need at least 112 calls; their exact optimum remains open.

The consequence is concrete: a standalone uncoding patch costs at least 112 paid calls while the removed eight H8 calls save at most 64 moment units. That route loses at least 48 units per FULL. This extends #4932's one-parallel-bank scope to active chronological calls; it does not extend that predecessor's per-channel grade-two conclusion.

The model uses same-width fresh fields and a fixed finite projective monomial group, with the same one-column map lifted across every digit column. The lower bounds hold at sufficiently large widths. Already-paid native observations, nonuniform per-column address operations, compressed sandwiches and full-controller admission remain outside scope. No improved κ or global caller impossibility is claimed.

Validation on publication commit 0e8a13c251a9be1029114cda8b4194f05f3a23ec:

• CPython 3.12.15 registered normal/optimized replay and separate binary checker pass.
• Exact source checks cover all 12,544 port pairs, 5,824 supported code terms, mode-factor/Gram identities and four rejection controls.
• Committed-byte/source-history preflight passes for all 17 package/input files.
• Official committed export integrity check and both optimized direct checkers pass without Git.
• Separate AI-agent proof reviews are included; no independent human or formal-kernel acceptance is claimed.

Publication branch starts from refreshed canonical main 20a1e31bc6009b9b3f4d084bbaadc6350e34fcba; frozen scientific sources remain pinned to 4c0a359d88f22f05189d51f48387750b801a45d4. Latest common-code moving-stage and coherent fresh-D work was reviewed and its distinct scope retained. Claim: #4930 (comment).

Credit Dugongue/#4929, Dillon/#4933, Chris R/Clear1nsight/#4932, the passive saved-query clarification, Cinna Davis, Cao Yifan, zxv, and the retained native/compiler and rank-cut lineage. Prepared for Dillon with substantial OpenAI Codex assistance. Selected frontier and predecessor files are unchanged.

[dillona]

Hosted status for 0e8a13c251a9be1029114cda8b4194f05f3a23ec: the planning job failed in the shared onboarding test suite, so the package replay job was skipped. This is not a hosted scientific replay pass.

The log reports eight failures in two shared test modules:

• test_research_descendant_source_roles: the existing isolated-descendant-bytecode input pin for scripts/check_research_commit.py rejects the current main version before the intended fixture checks. On base 20a1e31bc6009b9b3f4d084bbaadc6350e34fcba, the pin is ecf206a82bf1f4e748efb6cc639919582050d173b3537ec12f7265a4f7b55ead, while the base file hashes to 28fed1e156fe8601633ed453f682a158cb15e94981e6c89ed96ae72ffc9d0c6f.
• test_report_survives_a_cp1252_stdout: its temporary fixture copies only find_overlap.py, which now imports research_package. I reproduced the identical ModuleNotFoundError using only the file from that base commit and the fixture's Python 3.12 command, without this PR's package.

This PR changes only its new research directory. Its registered normal/optimized/independent checks, committed source-history/byte checks and Git-free official exported replay all pass locally on CPython 3.12.15. The shared fixture/source-role repair and a subsequent hosted replay remain outstanding; no shared infrastructure or historical pins were changed here.