Reproducibility archive for "Closing the per-field reduction: collar contraction, the stretched road, and a certified minimax for the complex sin²-algorithm"

Version 6 (archive sealed 2026-08-26, published 2026-08-27). SHA-256 of archive_paper3a_v6.zip: 6f0e4f9a7f82f47f20c682ddf07cdc4d06a42f564f193f1ceabac5ad161a34c0. Version map: the preprint pins archive v5 by its printed SHA-256 (3bd11bdf...19f175) — both the v2 PDF (80 pages, 2 August 2026, SHA-256 8196f501025bf5e7b7a3b83a732a5fcbdb83f2d7f56c0776216529189dbe0709) and Version 3 (4 September 2026, 75 pages, https://doi.org/10.5281/zenodo.22285098 , SHA-256 07df3a1fd6e4497cc71622cb13d967f9e4b2595a2a60a991b058aae41cd93386), a typographical revision, two errata and wording corrections; v5 remains available under this concept DOI. Note on counts: this archive ships the same 100-path MANIFEST.sha256 as v5; only eight file digests differ. Reproducibility archive for Closing the per-field reduction: collar contraction, the stretched road, and a certified minimax for the complex sin2-algorithm (direct half). Version 6 hardens the certification machinery after an internal pre-exposure audit of the published v5 objects. No mathematical certificate changed and no audit finding invalidated a published result: the corrections make the verifiers able to fail on classes of alteration that were previously accepted, and align the documentation with the measured behaviour of the full replay. Changes with respect to v5: - The wrapper's manifest control now fails closed, before any block, on any precompiled Python bytecode found in the tree, and every replay runs with its bytecode cache redirected outside the archive. The certified replay assumes a fresh extraction of this archive.- The collar recollement certificate now pins the shared collar width d_* = 1/10000 across every collar module (live attribute comparisons plus exact source anchors), fail-closed in both directions; the former self-referential equality in the q-wing module is replaced by independent pinned checks.- The lever completion-of-square identity now anchors its constant term (the centre-energy value) against an independently constructed comparison.- The retained collar sampling cross-check now exits nonzero, without its completion marker, if any sampled state is left unassigned.- Documentation: the complete --heavy-r --full-r path peaks at about 2.03 GiB of resident memory (more than 2 GiB; a 2-GiB machine cannot reach VERIFY_ALL_OK), and "portable" means CPython plus the packages in requirements.txt, not the standard library alone.- The LaTeX working source in the sources package carries the two journal-revision corrections: the marker of problem4/regfree_state_count_certificate.py is REGFREE_FINITE_REGRESSION_OK (the previously cited UNIFQUANT2_OK does not exist in the archive), and the Methods statement now excepts the analytic import listed in the owner table (namely the u_3 owner) and the companion-paper interfaces, which are proved in the companion papers and are not replayed from this archive; both corrections are published in Version 3 of the preprint. Each verifier correction was re-verified by the audit's paired negative control (the controlled variant is now rejected for the stated reason), and the complete 37-block battery (verify_all.py --heavy-r --full-r: 33 default blocks, the three --heavy-r blocks and the --full-r block, matching the 37 registered verifiers) was replayed to VERIFY_ALL_OK on a fresh extraction of the sealed v6 archive. Files: archive_paper3a_v6.zip (reproducibility archive), paper3a_sources_v6.zip (LaTeX sources and figure generators; the working source file paper3a_v2_work.tex is the journal-revision working copy; the authoritative text is Version 3 of the preprint (https://doi.org/10.5281/zenodo.22285098 ), whose sources are the separate package v7; the v2 preprint remains available unchanged).