Independently audit the P104 tail-prefix rank certificates

[RossRichardson]

PR #2358's P104 tail-prefix dimension certificates now have a separately implemented exact audit. Fraction-valued sparse rows, canonical binary elimination and an every-cut scan corroborate all 4,059 cuts, 280 independent input columns and 110,814 target/snapshot pairs. Exact compression functionals and 18/46 source-bound contacts reproduce correction ranks 99/98, including the explicit cut-2754 rank-two residual.

This adds verification evidence to AAAmirinu's result, not new minima. The parent inspected predecessor source before implementing the distinct interpreter; separate AI-assisted mathematical and code review checks the rank reduction, nuisance projection, exact reuse and compression arguments. Certificate origin, immutable sources, predecessor attribution and a written proof audit are retained. The original certificate is copied byte-for-byte; no predecessor Python code is imported.

Validation: optimized CPython 3.12 registered replay and all eight control groups pass after exact-zero hardening; a Git-free optimized CPython 3.14 export passes with all file contents unchanged. Controls cover nonzero compensation/cleanup faults, wrong/future/self contacts, scalar/private-frame faults, invalid lower generators and omitted compression premises; equivalent bases/contact order pass. Source-pin mutation and occupied-output checks reject. Selected-record integrity and staged whitespace pass. Hosted CI is separate.

python3.12 scripts/research_package.py --package research/p104-prefix-tail-independent-audit

Scope is the rational pooled unchanged-trace model: no chronological acquisition, native Gaussian regeneration, paid transport bound, owner map/moment, precision/all-size admission, selected cold replay, supplier or new kappa. Pooling snapshots at different times does not make a physical correction free. The active 41-contact chronological lift is a different obligation.

Base: freshly checked canonical main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, priority revision 4. Audited #2358 head e8950db523c9682518049a42cfc28642224292ff; selected files unchanged. Canonical claim. This is one manually launched saved-prompt trial; it does not test scheduler firing.

Credit AAAmirinu/#2358/#2345; Gonzalo Mier/#2173; Gareth Davidson/#2251; @josecarlosgonzalezroque-design's late-prefix handoff; Rohan Arun/#1972; Chafik Boukhalfa/#1991; Adam Gall/#1940; Josie Elliston/#1071; Ali Romman/#1125 and retained lineage. Prepared for Ross with substantial OpenAI Codex assistance, including separate exact-circuit specialist proof/code review. No independent human or formal-kernel review. Apache-2.0.