Research: batch JP8 compensation, saving 80,150 / 81,464 primitives (no new kappa)

[Varun-Mittal-1001]

Change and mathematical scope

The selected #626 JP8 program independently prepares coefficients for 8,800 frame-zero compensation additions. Their source and destination sets are disjoint, so the additions commute. Grouping them by source and using one signed quarter-grid coefficient route reduces the complete local real-component primitive bounds from 257,232 to 177,082 forward and 257,280 to 175,816 backward: savings of 80,150 / 81,464.

Status: finite research construction / complex-supplier circuit improvement; no new κ. Every source read, signed accumulation, scale and initial/final clear is charged. Both exact full-local prefix envelopes remain unchanged at 64/101 bits within the inherited 192-bit allowance. The 492 helpers, all recursive children and their paid profile are unchanged. The private coefficient stream clears at source-group boundaries rather than every original macro; production integration and refreshed global admission remain separate obligations.

Only research/jp8-compensation-batching/ is added. It includes the executable primitive panels, complete group/route certificate, proof and cost note, source pins, separate integer/dyadic checker, unchanged native-checker harness, validation receipts and attribution. upstream/, existing generators and the selected frontier are untouched.

Validation

python3 -I -B research/jp8-compensation-batching/verify.py

Expected: PASS_JP8_COMPENSATION_BATCHING additions=8800 saved=80150,81464 columns=716+9 controls=10.

• Full original/replacement scalar primitive projections in both orientations on 716 native plus nine arbitrary private input columns; exact endpoint, dirty-helper restoration, private clearing, cost and prefix checks.
• Separate integer/dyadic panel checker on 717 independent columns, without importing construction code; twelve further negative controls. Its committed receipt reproduces exactly.
• Unchanged inherited native emitter and scalar/frame/wrapper auditor on the reordered canonical panel: both orientations, all 716 columns and 25 existing controls pass. Native backward uses the dual of the forward permutation; the primitive backward panel uses its separately checked source permutation. This is native-order compatibility, not global admission of the new lowering.
• Linux CPython 3.11.13; normal/optimized execution, minimal relocated checkout with spaces, invocation outside the repository, alternate gzip compression, assertion-enabled native subprocesses under PYTHONOPTIMIZE=2, output-preservation, make selected-record-check, Ruff and staged whitespace checks pass.
• Full repository verification, cold supplier rebuilding, global hosting/invoice/outer admission, Lean and remote CI were not run locally. The inherited all-size/compiler, common-label/mutable-basis, whole-call hosting, precision/tape/recovery, bit and analytic assumptions remain. Finite replay is not a proof of the complete multiplication theorem.

Provenance and coordination

Exact base: 66849e1475c930fd7732b1a0067ffee7e1e95f60; selected #626: 660de893d27bb7abd1d5b01b1640207528a5c851; #622 construction: f69477f4f6d3c7eb1f740b166881bffed4ffa9f6. Both full inherited payloads match the preserved #626 SPEC. OS-random assignment seed 27 selected the complex track and JP8 by parity. The complete bounded search covers 266 forward / 112 backward source groups and all 15 nonempty magnitude subsets on the signed quarter grid; there is no general circuit-optimality claim.

No pinned coordination issue exists, so the claim is local. The pre-submission refresh covered 607 open PRs through #788 and confirmed unchanged upstream main. Concurrent #722/#739/#755/#769 are credited for order-preserving coefficient sharing/rescaling; #749 changes/restores the compensation target basis; #787 composes within-macro lowering with cleanup sharing; #786 factors center quarters and audits its macro overhead. This contribution changes the commuting compensation order. No pending code is imported, no priority claim is made for coefficient sharing itself, and savings are not asserted additive.

Credit to Jeremy Kothe/AtomScout and the #622 lineage, Chafik Boukhalfa's #626 package, icekylinx, DreamingOfClouds, Dominik Scholz, jon314159, Dugongue, Rohan Arun, eumemic and retained predecessors. Prepared for @Varun-Mittal-1001 with substantial OpenAI Codex assistance, including separate AI mathematical, implementation and adversarial checks; no independent human review is claimed. New files are Apache-2.0.

[Mister-Meeseeks]

100-PR batch review at f9383f324ee86793c5ae7b73f8e39d5827a8aa52.

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

Reviewed scope: JP8 signed-quarter compensation batching: complete local rows, scratch and prefixes in both directions; recursive profile unchanged.

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.