Tooling: fuse exact dyadic updates for a 1.45–1.50x faster PR799 audit

[bitplane]

The PR799 independent scalar audit repeatedly constructs exact rationals for y + q*x. This opt-in adapter fuses dyadic updates into one normalization while retaining ordinary Fraction rows and every original native check. Verification tooling; no new κ. Original sources and default commands are unchanged.

Five alternating fresh-process pairs per interpreter measured 1.45–1.50x faster complete audits on the same fixed word:

Linux CPython	Original wall median	Fused wall median	Ratio
3.12.3	0.559843 s	0.374207 s	1.496x
3.14.4	0.501354 s	0.345660 s	1.450x

Every timed receipt matches the retained maintainer receipt in every field except elapsed time. Imports, source admission, parent-interface binding, native checks, output and serialization are included. Raw wall/CPU samples, source and implementation hashes, environment and pilot evidence are committed. Timing is descriptive; this is one audit stage, not a whole-frontier or multiplication speedup.

Only research/scalar-fused-dyadic/ is added: exact kernel, source-bound adapter, proof, full differential verifier, tests, raw benchmark evidence, pins and local claim.

python3 -B -I research/scalar-fused-dyadic/verify.py
python3 -B -I research/scalar-fused-dyadic/test_kernel.py
python3 -B -I research/scalar-fused-dyadic/benchmark.py --rounds 5 --output /tmp/new-scalar-benchmark.json

Validation includes:

• All 728 arbitrary native columns in both orientations, compensation reconstruction, native frames, entrance/width/profile checks, parent binding and helper row/column cones.
• Exact target-row/source comparison at all 70,765 actual updates, plus 810 arithmetic cases and 10,240 additional complete state prefixes. No sampling or floating arithmetic replaces the original checks.
• 15 changed-word controls reject with identical exception types/messages; three invalid-gate cases and nine tests pass. Two deliberately faulty optimizers fail the complete audit.
• Python 3.12.3 and 3.14.4, optimized/no-site runs, minimal relocation, source/receipt corruption, input/archive repackaging and output preservation pass. Selected-record and staged whitespace checks pass.

Exact base: f567884aeb8851b6e794af1184119a3bfbe1e631. OS seed 93, verification engineering. Final overlap refresh covered 641 open PRs through #844; upstream was unchanged. No sprint coordination issue existed, so the claim is local. #824's historical stage-pair optimization and the earlier interval, determinant, owner-mask and archive tools concern different kernels; no pending implementation is imported.

Credit icekylinx for the original independent auditor, Jeremy Kothe for retaining it in #799, and Chafik Boukhalfa, DreamingOfClouds, Rohan Arun, eumemic and the retained lineage. Prepared for Gareth Davidson (@bitplane) with substantial OpenAI Codex assistance, Apache-2.0. No independent human review is claimed.

No full repository suite, supplier cold rebuild, phase/wrapper replay, global admission/assembly, Lean or hosted CI pass is claimed for this optional adapter. All inherited compiler, realization, hosting, recursive, precision/tape and analytic obligations remain. The selected frontier is unchanged; finite verification is not a full multiplication theorem proof.