[petalas]

Summary

PR3148's pinned data
  -> independent time-major replay of all 2,717,280 roles
  -> integer-square-root moment enclosures and all-pattern dual
  -> corroborated allocation gain and attained fixed-inventory cap

Adds research/coupled-cap-independent/. No discrepancy was found in PR3148's 145-row exchange. The actual T74-to-T740 root gain is strictly positive and agrees with the reported approximately 1.4052734766776475e-9 improvement. All 2,551 patterns are certified, with 12 tight patterns, and every literal bin is tight. This rules out further packing-only gains for the supplied fixed profile/inventory, not other words or interfaces.

Audited head: f2927a4b8afcf0701a6b945e14f30de1d94f5176. Base main: 997e6a9ca7b4f95d50e366f61412fd640d72a88a, priority revision 6. Canonical claim and finding.

Credit Jeremy Diviney/PR3148 for the construction, dcharb78/PR3133 for the words/inventory/fill, Gonzalo Mier/PR3143 for the comparator, and the retained supplier, allocation and cap lineage. Notices and exact upstream blob identities are preserved. The first commit pins byte-identical data; the second adds the independent audit and reachable-ancestor replay pins.

Evidence

• Before: the new changed-inventory journal and cap used the retained owner/lift/series-arithmetic implementations. This audit imports none of them.
• After: all 2,717,280 roles and 1,491,840 bins pass source-window, physical-family, disjointness and compact-prefix checks. A different exact enclosure method certifies both literal allocation roots and the 2,539 strict pattern inequalities.
• Registered Python 3.12 replay passes, including 14 behavioral groups, optimized execution, exact-power controls and a Git-free read-only export with source-corruption rejection. Direct Python 3.14 replay, Ruff and 16 onboarding tests pass.
• Publication byte preflight passes all 23 package/input files. The primary local replay took 6.6 seconds with approximately 195 MiB peak RSS. Hosted CI is separate.

python3.12 -B scripts/research_package.py --package research/coupled-cap-independent

The supplied ordered widths and local paid profile remain inherited inputs. Native operators, producer algebra, the full physical/prefix/outer release, the bit supplier and all-size/analytic obligations are not independently checked here. No new construction, selected-frontier update or independent human/proof-kernel review is claimed. Substantial automated assistance was used in analysis, implementation, local review and writing.

Merge Danger

Door: two-way. New isolated research package only.

Blast radius: research. No shared checker, selected record, frozen predecessor or workflow is changed.