[aliromman]

Thirteen paid native conversions retain current source values for later scalar contacts on exact PR5751. The new word changes seventeen helper targets and removes thirteen other transports. Conditional κ rises from 0.00311301600 to 0.00312192127, a gain of 0.00000890527 (0.2860657%).

The paid-call and target counts stay fixed. The signed delta is 8H4 + 16H5 + 6H7 - 12H1 - H6 - 17H8. An exact majorization certificate proves its gain for every power exponent 0<t<1. The complete rational calculation encloses the raw crossing in (0.00312192130450, 0.00312192130451).

The package is research/p104-retained-native-copy. It provides a fixed constructor and plan, a literal word, source-version and ownership witnesses, a complete proof, rejection controls and an authenticated predecessor export. It uses 767 arbitrary public inputs, 22 owned labels, 24 owned lifetimes and peak 12 live owned fields. All 13 new COPY/query/use/CLEAR paths are charged. The complete caller retains 66,834 paid jobs, 13,390 targets and rank surplus 760.

Validation: primary, registered normal and optimized replays pass; all 789 input columns, 2,925 native maps/phases, the complete caller, two rational interval engines, compiler, 54 strict outer predicates and four functional checks pass. Nine word/caller/invoice controls and the native phase control are rejected. Metadata/pins, committed bytes, source ancestry and export integrity pass. A separate read-only, Git-free -I -O -B replay with empty PATH passed in 104.635 seconds; every file byte and mode remained unchanged. The full-history preflight was not run in this shallow checkout; both declared top-level source ancestries were checked directly.

Replay: python3.12 -B scripts/research_package.py --package research/p104-retained-native-copy (standard library, two registered modes, 600 seconds per mode).

The five inherited framework assumptions remain. No large uniform word or rounded multiplier is materialized. Independent acceptance remains pending; the selected frontier is unchanged. Hosted planning CI failed before scientific replay. The replay job was skipped. Its eight failures and two errors in test_research_current_assignment.py reproduce locally. The test, both picker scripts and the task catalogue are byte-identical to main; the tests still expect the old command mappings after revision8. No hosted scientific pass is claimed.

Source: PR5751 at 8483f9a679d06c1e270141dbc686a7e76e7a7522, preserved in this branch's authenticated archive. The branch starts from canonical main de5ec2e7a6074068452f45b1a191b5c72832df4a, priorities revision 8. Discussion5645 records the separate retained-copy scope. The R22 block composition is not included.

Credit Ali Romman's PR5751 and P104 work, flyingviper PR5718, Joel Pulikkan / GamingPuzzled PR5739 for the dimension-four basis, Cinna Davis's controller, and all native/compiler/BIT contributors retained in NOTICE. New construction, search, verification and proof: Ali Romman with substantial OpenAI Codex assistance. No subagents. Apache-2.0.

Publication head: 7d6ec0cdef31d0c37b7d2563e709979caa0d5eae. The 19-file export SHA256 is e47cd61d6542497287f5a69820484e92f7116aaf76663969fe59a1d85fffe497.

Proof · Exact result