[aliromman]

Three pending sums now share native conversions before their destinations are read. The complete conditional κ increases from PR5786's 0.00312265044 to 0.00312336057, a gain of 0.00000071013 (0.0227413%). This is a modest gain. The selected frontier is unchanged.

The fixed constructor uses three owned buffers on exact PR5786 at 6b936acd68a7e0f9062037900f0fe9bc4b6fdd25. Three paid sum queries replace five source queries. It checks the actual destination read barriers and retains every arbitrary public input, helper target, native map and charge. The local signed gain is 2*((2H1-H2)+(2H7-H6-H8)), which is strictly positive for every 0<t<1. The complete p16/a7 caller has 66,870 paid calls, 13,444 targets and rank surplus 760. Both interval engines certify the full price and compiler completion; the strict raw crossing is (0.00312336059973, 0.00312336059974).

The contribution preserves all 1,704 protected events and public boundaries from the original R22 and PR5796 contact regions. All 23 retained source COPY operations keep their exact decoded values. It excludes the provisional role 92 accumulation, which touched the separate block investigation. PR5795/PR5796 composition remains separate. This result exceeds PR5795's published κ by 0.00000058744; no combined result is claimed. Discussion5125's concurrent transposed-copy route uses the same duality and is credited. This PR claims these three fixed forward-supplier groups and their complete price, not the elementary accumulation identity.

Proof, fixed plan, complete result, source pins, validation limits, search coverage.

Validation on macOS arm64 / CPython 3.12.13:

• The primary replay passes with warnings treated as errors. It checks all 798 input columns, 636,804 endpoint coefficients, a separate dense transpose, 2,924 native maps with absolute phases, 1,124 helper normalizers, 208 main adapters, the physical caller, complete invoice and all 54 strict plus four functional compiler checks. All declared mutation controls reject.
• Registered normal and optimized replays pass. Metadata, source pins, committed bytes and executable modes pass. Direct canonical-main and archive-commit ancestry pass. The deterministic 20-file export and its declared-input check pass.
• A separate sealed export replay passed in 99.853 seconds with Python -I -O -B, empty PATH, no Git checkout and read-only payload files. Every payload byte and mode stayed fixed. Export SHA256: 3e83941cd643f1c6b9288fa610a0f607dd94dc21bcbba6436295ad09ff1f81fb. Replay log SHA256: cf44b11cc951893cd2598bc372b7ea4a7a02950f9b6d063107281fa69e43d55c.

Reproduce with Python 3.12 and the standard library:

python3.12 -B scripts/research_package.py --package research/p104-accumulated-native-updates

Exact publication head: 9bab1ca917ec4d1a4e60f4eb8c81d7ef92908a62. Canonical base: de5ec2e7a6074068452f45b1a191b5c72832df4a, priority revision 8. Canonical investigation and scope.

The five inherited framework assumptions remain: native_all_columns, owned_erasure, compiler, precision_tape, and analytic_recovery. The large uniform word and rounded multiplier are not materialized. These are author checks; independent acceptance and hosted scientific replay are not claimed. The optional full-history preflight is not claimed in this shallow checkout. No merge, closure or frontier update is requested.

Prepared by Ali Romman with substantial OpenAI Codex assistance. No subagents. Apache-2.0. Credit PR5786, PR5774, PR5751, flyingviper's PR5718, Joel Pulikkan / GamingPuzzled's R22 work and every retained P104/native/helper/controller/compiler/BIT source in NOTICE. The authenticated predecessor archive retains their exact source bytes and notices.