Research: share JP8 cleanup coefficient work, saving 19,688 primitives (no new kappa)

[ellipsi2]

Change and mathematical scope

The selected JP8 cleanup repeatedly prepares the same non-unit dyadic coefficient for consecutive reads of an unchanged source. This finite circuit result shares that preparation across 999 runs / 3,022 additions, using the existing coefficient-work stream and retaining the original target-update order. It saves 19,688 real-component primitives: the full forward scalar count decreases from 257,232 to 237,544 (7.65%). Backward has no eligible run and remains at 257,280. Status: research note / finite complex-supplier circuit improvement; no new kappa.

Only research/complex-jp8-cleanup-fusion/ is added. It contains the executable primitive lowering, every run's exact certificate, a cost/prefix argument, source pins, independent arithmetic checks, scoped AI review, local sprint claim, and validation record. No existing supplier, emitter, verifier, selected frontier, or upstream/ file changes.

The local interface change is explicit: coefficient work is cleared at each new fanout-block boundary, rather than after each original scalar macro. The block preserves arbitrary live inputs and clears arbitrary incoming private work. All source reads, Horner operations, signed accumulations, and clearing are charged. Helper count, paid frame/child profile, wrappers, restoration map, and exact prefix envelopes remain unchanged. Production-emitter integration and global invoice/admission regeneration remain outside this contribution; inherited compiler, realization, hosting, precision/tape/recovery, bit-supplier, and analytic assumptions remain. No full multiplication theorem or practical runtime speedup is claimed.

Base: 66849e1475c930fd7732b1a0067ffee7e1e95f60. Selected #626 source: 660de893d27bb7abd1d5b01b1640207528a5c851; reviewed #622: f69477f4f6d3c7eb1f740b166881bffed4ffa9f6; retained #543: 2065fe4f72a914d907f1046603f5a36f17018535. Consumed files and complete gzip payloads have SHA-256 pins; native and event payloads are also checked against #626's SPEC.

OS seed 31 assigned the complex track and selected JP8 by parity. A complete initial scan found no equal directed additions before their first commuting obstruction; the single pivot was shared coefficient preparation in that same cleanup. Both orientation scans are exhaustive for maximal contiguous non-unit common-source/common-magnitude runs, with no sampling, timeout, or optimality claim.

No pinned sprint issue exists, so the claim is local. Full-queue overlap review included 536 open PRs through #717. Related #696 lowers individual scalar macros; this shares work across macros and retains the original Horner preparation. Savings are relative to the selected base and are not asserted additive to #696 without a composed recount. #681 changes a raw factory word; #698 concerns CK B frames; #708/#710 change F2 bit circuits. No pending patch is imported.

The immediate pre-submission refresh reaches #721. Concurrent #719 is credited for its distinct CK cleanup rank-two obstruction: it excludes certain scalar-word replacements, while this JP8 construction preserves the scalar word and shares private coefficient preparation. These results are compatible and neither depends on the other.

Validation

python -B research/complex-jp8-cleanup-fusion/verify.py

Expected:

PASS_JP8_SHARED_COEFFICIENT_CLEANUP groups=999 saved=19688 columns=716 orientations=2 controls=10

• Native Windows, CPython 3.11.15, standard library only. All 716 arbitrary input columns replay in both orientations. Changed blocks execute explicit primitives; unchanged macros use exact rational semantics. Both inherited rational prefix envelopes match exactly.
• Ten arithmetic controls include missing work clearing, wrong signs/coefficient preparation, aliasing, and three endpoint-neutral prefix excursions.
• python -B research/complex-jp8-cleanup-fusion/test_verify.py: seven tests, including all 999 block maps in a separate flat-vector column interpreter, independent grouping/pricing, a noncommuting inverse-transpose witness, emitted-frame and ended-helper mutations, and omitted paid child.
• Isolated optimized Python, minimal copied checkout in a path with spaces, alternate gzip compression, source/receipt corruption, and existing-output preservation checks pass.
• python -B scripts/verify_selected_result.py and staged whitespace checks pass. Full make verify, selected cold replay, global admission, Lean, and remote Linux CI were not run or claimed.

Credit: icekylinx's direct-delivery/factory and emitter lineage; Jeremy Kothe/AtomScout's #622 and JP8 predecessors; DreamingOfClouds' cold-helper/native checks; Chafik Boukhalfa's #626 package; eumemic's #543; and all retained predecessor notices. Prepared for @ellipsi2 with substantial OpenAI Codex assistance, including one scoped AI reviewer. No independent human review is claimed. New work is Apache-2.0.

[Mister-Meeseeks]

100-PR batch review at a0766ab0fab379e1997fadc3ca157578d57986c1.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: JP8 shared-coefficient cleanup lowering: 19,688 forward primitives saved; complete rational rows, private scratch clearance and unchanged recursive events.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.