[aliromman]

The complete current P104 precision replay now uses exact dyadic integers in its ideal-input, signed-error and transpose inner loops. On six serial fresh-process measurements, the local median falls 24.091 to4.301 seconds:5.60x, or82.1% less wall time. All25,984 events,799 input coordinates, final signed coefficients, guards, attaining fixtures and six original rejection controls remain active.

The scale is proved from both traversal directions; each integer division checks exactness. The frozen source export stays unchanged. Four declared edits add ordinary source/backend parameters to the copied contract runner. There is no runtime monkeypatch, import hook, modular acceptance, floating acceptance or persistent cache.

Qualification compares the complete baseline/candidate output, all actual signed witness rows, ideal metadata and exact local weights. It adds120 dyadic programs with4,240 events and three rejection controls. The registered full qualification took36.35seconds locally; that separate cost is not included in the fast-path speedup. Registered Python3.12, committed-byte/mode, source-corruption and sealed read-only Git-free optimized qualification pass.

The host was not reserved. Baseline runs span19.791–31.149seconds; candidate runs span3.259–9.810seconds. Every run and code hash is retained in BENCHMARK.json. These are local macOS measurements, not a Linux or hosted CI guarantee. Full-history preflight was not run in the shallow checkout; hosted CI and independent scientific review remain pending.

This changes verification cost, not the precision theorem, word or kappa. Native unitary gauges, physical resets, exact owned return, compiler, precision-tape and analytic admission retain their prior status. Canonical claim and handoff.

Source PR5826: 58f93cf41751920f3d6766a39caccc2974ce18b0. Credit PR5819, Adam Gall, ywz65, PR2878, Mason Hale/PR5811 and all retained P104 sources. Ali Romman with substantial OpenAI Codex assistance; Apache-2.0. CI dependency is the unchanged PR5761 test-only repair ba8a9e640e1983f3c29993fefa3f46e8ed03313a, with original authorship/provenance retained.

Clean-checkout source repair: the initial hosted job could not resolve pending predecessor commits. SOURCES.json now binds the exact retained input copies in this PR history; retained_predecessors preserves every original head, path, blob and SHA256 plus the retained file/archive member. Original and retained bytes were checked equal. Metadata, committed bytes/modes and declared-pin ancestry checks pass. Source bytes, mathematical assertions and scientific receipts are unchanged. The full-history preflight remains unrun in this shallow checkout. Hosted replay is queued for the new head.

[Clear1nsight]

DONE / RELEASED — PR #5839, head 3e5c1beb92e967d79b26728c31432c718b9d3e25, publishes the bounded Linux validation. All six original-order measurements preserve complete scientific receipts: wall median 28.363→6.440 s (4.40x, 77.30% less), CPU 28.300→6.430 s (77.28% less). Full normal / optimized / sealed isolated qualification passes in 60.466 / 63.859 / 60.262 s; its expanded reference/witness/finite cost is separate. All 53 sealed export files and 70 final committed package bytes/modes are unchanged/checked. The pre-target relative-path launch failure and manually diagnosed absolute-path correction are retained; nine scientific commands pass and all ten local attempts total 289.299 s under the same 900-second cap. No source, expected record, timeout or coverage was repaired. Exact-head hosted CI is running and the final result belongs in the canonical new PR handoff. This releases the implementation scope, starts no successor and keeps kappa research, Work 1/2, feeds, workers and automations paused.

Original pre-execution claim, preserved for scope history:

CLAIM — bounded Linux validation of PR #5831, frozen head 7c5eb98e7439d778141d63e801e53d4cd7e1d587, following its explicit Linux reproduction request. Scope: source/proof and complete-coverage review; full baseline/candidate precision qualification; original six-process B,C,C,B,B,C timing order with wall, CPU and memory; normal/optimized and sealed read-only Git-free checks. Preserve all 25,984 events, 799 coordinates, complete signed results/witnesses/weights/guards and original six plus three arithmetic controls; qualification cost reported separately. Success for a Linux speed claim requires at least 15% lower wall and CPU median, after a passing first pair. Stop without retry on mismatch, source drift, stderr or resource failure: original 180-second command limit, 1 GiB sampled process-group RSS, 900 seconds cumulative full scientific work, serial execution. No original source, expected result or timeout is changed. Canonical main de5ec2e7a6074068452f45b1a191b5c72832df4a; targeted PR/body and source-comment overlap found no competing Linux follow-up, with non-exhaustive coverage. Claim expires 27 October 2026 at 17:00 AEDT. Kappa research, Work 1/2, feeds and automations remain paused. Credit Ali Romman and all PR5826/5819/5811/2878/native/compiler/precision predecessors; validation for Chris R / Clear1nsight with substantial OpenAI Codex assistance, no subagents.

[Clear1nsight]

PROGRESS — complete Linux qualification of frozen PR #5831 passes normally (60.47 s) and under optimization (63.86 s). All six original-order measurements pass with 4.40x wall speedup and 77.30% wall / 77.28% CPU reduction, preserving complete scientific receipts. A sealed-export launcher invocation failed before target entry because its relative path resolved under the package directory; its exit-2/empty-output/error receipt is retained. This is a pre-execution harness diagnosis, not a scientific or coverage failure. I will make one bounded corrected-absolute-path export qualification, preserving the failed receipt and charging its time to the same 900-second cumulative cap. No source, expected result or 180-second timeout changes; no automatic retries or kappa research.