[petalas]

Summary

A finite packing improvement to PR3018, using a doubled T=148 chronology and an 86-row journal that moves 50 helper roles. All exceptional [4,8,9,9,9,9]@48 bins disappear.

- [0.0017290342130642563095108, 0.0017290342130642563095109]
+ [0.0017290353460708281979174, 0.0017290353460708281979175]

The certified gain is 1.1330065718884065e-9 to 1.1330065718884067e-9, closing about 73.01% of PR3018's remaining fractional-packing gap. The exact complete delta is 4H1 + 4H3 - 4H5 - 2H14 + H32 < 0 for every 0 < b <= 1/100. The new empty capacity-32 bin stays paid.

Proof, journal, exact sources and reproduction. Canonical investigation.

This is a finite conditional moment improvement, not a new admitted kappa. Replica-dependent storage doubles; the finite physical namespace is 133,117, including all 91 private buffers. Full native/dirty restoration, physical backing/movement, precision, T-dependent constants, outer release and all-size/compiler admission remain unrun for T148. PR3021's complementary T74 restore/packing release does not bind these new artifacts. The selected frontier is unchanged.

Source PR2979 is pinned at be191346391661f4d35fc0d1508d0a6029572576; PR3018 at ca7113157d7c959fdfa9d7cd17ef92d3877070d0. Base canonical main d8e0f83f00191df62174b641e9fff055c357f641, priority revision 6. Credit Marius-Paul Dumitrean, Nick Petalas/PR3018, Mason Hale, Jeremy Kothe, Jeremy Diviney, Yiyang Lu, Tekalign, Ross Richardson/PR2987/3010, Dugongue, icekylinx, themilkman216 and the retained lineage. Assistance and review limits are recorded in TASK.json.

Evidence

• Before: exact doubled PR3018 invoice, four exceptional bins, two paid H14 remainders.
• After: all 543,456 roles and 298,368 bins pass full source-window/family/mask/prefix checks. The retained owner engine agrees; a separate sorted-interval recount reproduces every idle rank and all twelve target patterns. One paid H32 replaces the two H14 remainders as part of the complete delta.
• Registered CPython 3.12.15 replay, nine rejecting controls, four behavioral groups, optimized CPython 3.12/3.14 execution, warning-as-error checks, Ruff and exact committed-byte verification all pass locally. Hosted CI is separate.
• Reproduce: python3.12 -B scripts/research_package.py --package research/rank7-doubled-residual-exchange.

Next useful work is source-bound admission of this literal T148 order/fill, with every T-dependent physical constant rederived. No further search or integral-optimality claim is made.

Merge Danger

Door: two-way

Blast Radius: isolated

One additive opt-in research package. No frozen source, shared checker, selected result or other PR disposition changes.