[petalas]

Summary

PR3042 six carrier/pivot skips + PR3030/3024 exact T740 allocation
  -> all 2,717,280 roles and 91 private buffers charged
  -> fresh source/native/admission/modern-outer release
  -> conditional kappa = 0.001729737950064278551242 - 10^-46 - 10^-70

Task selected-packing, priority revision 6, scope changed from the admitted selected word to this bounded pending-word composition. It closes the new word's remaining packing gap without another search.

Concurrent PR3050 supplies a complementary substitution audit without a fresh native/admission/outer run; it is cited in the package.

The packing-only gain over pinned PR3042 is approximately 1.78367636e-9. This attains the exact fixed-inventory additive-bin cap for its six-skip words. Only research/rank7-t740-carrier-skips/ is added; the selected result is unchanged.

Canonical investigation. PR3042 is pinned at 84cc406ad91ec53fff329a11a445d7effe564a77, PR3030 at 78e77d820393c6b8648b47af219e991edfcc54e6, and main at d8e0f83f00191df62174b641e9fff055c357f641, priority revision 6.

Evidence

• Before: PR3042 releases its six-skip words on the original T74 fill at 0.001729736166387915986812 - 10^-46 - 10^-70. After: the exact T740 composition releases the endpoint above. Its complete paid profile has mass 48,165,120; no helper or idle cost is omitted.
• An exact rational derivative proof shows the complete packing delta 60 H1 + 57 H3 - 75 H5 + 3 H48 is negative throughout 0 < b <= 1/100. All 2,551 cap patterns are recertified at the changed root, with 12 tight and 2,539 strict.
• The fresh full chain passed all 13 stages in 1,749 seconds, including its retained bootstrap, both edited native providers, the complete literal admission and both modern outer checks. The actual invoice charges 663,549 physical families and a 1,479-bit global prefix under 2,048 bits; the local prefix is 95 bits under its 128-bit reserve.
• Fourteen behavioral groups reject dirty-action/cleanup/width/CLI/source/capacity/paid-mass faults, a real stale R1 invoice, stale native artifacts and understated prefix/stock. Six additional stored-release mutants also reject. The standard-library registered replay, optimized test run, strict Ruff checks and source compilation pass. Publication bytes are checked against the final commit below.

Registered finite replay, source pins and stored fresh-release relations:

uv run --python 3.12 --no-project python -B scripts/research_package.py \
  --package research/rank7-t740-carrier-skips

Source, proof, evidence and reproduction. All 95 committed package/input byte comparisons pass at that exact commit. Canonical main was fetched again immediately before publication and remains the pinned baseline. The README supplies the separate full replay, exact Git exports, environment pins, sources, engineering bounds and open next question. These checks reuse credited source arithmetic/geometry; they are not a new independent implementation or a full multiplication theorem.

Credit dcharb78 for PR3042's six skips, Jeremy Kothe for PR3015's edit stage/R1, Marius-Paul Dumitrean for PR2979, Nick Petalas for PR3018/3024/3030, and the retained native/geometry/owner lineage. Substantial automated assistance was used; no independent human or proof-kernel review is claimed.

Open next questions

Independently replay the exact combined source/native/owner/outer construction before admission. Start from the pinned exports and compare the words, literal fill, invoice and release receipt; the documented environment and roughly 30-minute full replay are the practical blockers. Inherited theorem premises remain separate.

Merge Danger

Door: two-way.

Blast Radius: isolated.

One additive research package. Physical backing/movement, all-size/compiler realization, precision/tape, caller, analytic-transfer and universal block-energy hypotheses remain inherited. No selected-frontier or other PR disposition changes.