DONE: Windows selected PR1094 successor replay (PR2336)

[antunishdPursuit]

DONE: native Windows successor evaluation, public artifact PR2336, current head 217ba488beeb20d65a3f677df46de9793b954c50. Base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, priority revision 4. PR2316's two in-memory path-key conversions permit all eleven strict stages on Windows / CPython 3.12.14. All 105 outputs compare: 88 byte-identical, 11 CRLF-only logs, one reordered invoice, six relocated input-path keys, and checked dependent hash links/timings. Sixteen corruptions reject. The branch now retains #2055's historical CI source-compatibility repair; all 64 fork jobs across eight workflows pass. Original selected Windows entry point still rejects inventory admission; assembly bytes differ through the checked invoice hash. No cold ancestry, new kappa or theorem claim. Credit PR2316, Rohan Arun, Jeremy Kothe, Douglas Colkitt, JosieElliston and the retained lineage; substantial OpenAI Codex assistance, self-reviewed only.

Complete 52-job Ubuntu matrix; scoped exact tooling/provenance audit. Local reconciliation and 43 targeted regression tests also pass. The package's actual successor replay remains evidenced on native Windows; its registered Linux replay is queued upstream and is not supplied by the historical matrix. Full historical Windows producer/tests remain platform-limited by the missing C++ compiler and legacy POSIX receipt-path keys, without any weakened equality checks.

The complementary PR2101 comparator adapter claim below remains separate. This package does not edit or reuse that comparator. Source pins, receipt and open next questions are in research/selected1094-windows-successor/: full cold Windows ancestry and maintainer adoption of a versioned portable entry point. Selected sources and frontier are unchanged; no PRs were merged or closed.

[AAAmirinu]

DONE / scope released: the complementary exact six-input-role adapter is published as PR2344, head b599a781bfa8a8fb09a8dea7fa924edb353c21cb. Canonical finding and full scope: #1472 (comment) . Normal/optimized registered replay passed 6 positives and 176 negatives each. These are constructed relocations of archived outputs; your PR2336 native execution remains separate. The adapter retains PR2101's strict byte-exact inner-log comparison, so your reported CRLF differences require a separately reviewed composed log policy. Credit all predecessor work in the package; substantial OpenAI Codex assistance for @AAAmirinu.