Research: save 48,400 JP8 primitives with a restored compensation basis (no new kappa)

[theon199]

The selected #626/#622 JP8 supplier's frame-zero compensation contains 8,800 additions. Factoring its rational target rows through 56 disjoint, restored target pairs reduces that to 4,596 additions, including every preparation and restoration gate. The complete emitted scalar program saves 48,400 real-component primitives per orientation under the unchanged coefficient lowering.

Status: research construction / finite complex-supplier circuit improvement; no new κ.

Quantity	Original	Replacement
Full-program scalar macros, either orientation	22,693	18,489
Forward real scalar primitives	257,232	208,832
Backward real scalar primitives	257,280	208,880

The proof preserves arbitrary target and helper inputs through T^-1(Ty + TCh) = y + Ch. Helpers are read but not written during the replacement. Both target basis changes, coefficient preparation, source reads and clearing are charged. All non-scalar events, frames, 492 helpers, paid children, center operations and ordered wrappers are unchanged. Forward prefix growth is unchanged; the backward bound increases slightly while remaining at 101 bits, below the inherited 192-bit local allowance. No runtime speedup, global invoice or changed recurrence exponent is claimed.

Only research/complex-compensation-basis/ is added. It contains the written proof, exact recipe/certificate, source pins, bounded search, separate integer arithmetic replay, rejection controls, local claim and validation record.

python3 -I -B research/complex-compensation-basis/verify.py

Expected: PASS_COMPLEX_COMPENSATION_BASIS pairs=56 compensation=8800->4596 saved_real_primitives=48400 columns=716 controls=10+25 kappa=unchanged.

Verification passes on macOS arm64, CPython 3.14.5:

• All 716 input columns of both original and candidate scalar programs in both orientations; a separate Fraction boundary check; unchanged native emitter and frame/wrapper auditor on the changed word.
• Ten new controls, 25 inherited controls and six focused mathematical tests. Omitted preparation passes helper-only testing but fails arbitrary target-input testing.
• Exact reproduction of all 124,320 pair/coefficient scores and one seed-24 greedy disjoint pass; 3,095 profitable candidates. No optimality claim.
• Normal and optimized execution, minimal relocated checkout, alternate gzip packaging, corrupt source/recipe/certificate rejection, and assertion-enabled native subprocesses under an optimized environment.
• make selected-record-check, documentation links and staged whitespace pass.

Exact base: 66849e1. Selected #626: 660de893d27bb7abd1d5b01b1640207528a5c851; #622: f69477f4f6d3c7eb1f740b166881bffed4ffa9f6. Full payload hashes also bind the input word/programs to the selected SPEC. OS assignment seed 24, complex-supplier track. No pinned coordination issue exists; the claim is local. Final overlap inspection reaches #748; upstream main remains the exact base.

Credit to Jeremy Kothe, Chafik Boukhalfa, DreamingOfClouds, icekylinx, Dominik Scholz, Dugongue, eumemic and retained predecessors. Concurrent #713/#716 are credited for restored target bases in binary R420; this transfers that mechanism to rational JP8 with coefficient-dependent prices and cold-wrapper obligations. #696/#723/#734 lower individual coefficients; #722/#739 share cleanup coefficient work; #681 changes raw factory gates. No pending implementation is imported and no additive saving is asserted.

Prepared by Sricharan Vagvala (@theon199) with substantial OpenAI Codex assistance, Apache-2.0. Parent-only execution; no independent human review is claimed.

Production integration still requires a typed alternative-compensation construction step, refreshed bindings and global hosting/invoice/outer admission. All inherited compiler, realization, hosting, precision/tape/recovery and analytic assumptions remain. Full repository/cold supplier replay, Linux CI and Lean were not run. This is a finite circuit result, not a formal multiplication theorem. The selected frontier and upstream/ are untouched.

[Mister-Meeseeks]

100-PR batch review at d76c276a9f77df270982a61be7e53219316faf55.

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

Reviewed scope: Complex target-pair compensation basis: restored rational boundary maps and both native orientations; 48,400 local primitives saved each way.

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.