Research construction: rescale JP8 coefficient work across magnitudes (no new kappa)

[namredips]

Change and mathematical scope

The selected #626 JP8 cleanup prepares each dyadic coefficient independently. This finite circuit contribution reuses its private work stream across consecutive different coefficient magnitudes related by powers of two, preserving source values and target-update order. It saves 10,338 additional real-component primitives beyond the equal-magnitude sharing in #722.

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

Complete scalar program	Original	Equal-magnitude sharing (#722 method)	Rescaling
JP8 forward	257,232	237,544	227,206
JP8 inverse-transpose	257,280	257,280	257,280

The 30,026 total saving includes #722's 19,688; only the additional 10,338 is new here. All source reads, Horner preparation, doubling/halving, signed accumulation and clearing are charged. The original exact prefix envelopes and paid recursive profiles are unchanged. This is an operation count, not a runtime benchmark.

Only research/complex-coefficient-rescaling/ is added. It contains executable primitive circuits, all 1,015 block certificates, a written substitution/cost argument, five source pins, an independent DP and scalar-basis checker, negative controls and a local sprint claim. No pending implementation is imported.

The local interface requires clearing private work at block boundaries, replacing the selected emitter's per-macro clearing contract inside those blocks. Production-emitter integration and refreshed invoices remain future work. No additional helper, frame transition or recursive child is introduced. The selected supplier, upstream/, global ownership and frontier pointer are unchanged. Inherited compiler, realization, hosting, precision/tape/recovery, bit-supplier and analytic assumptions remain; this is not a formal multiplication theorem.

Validation

python3 -I -B research/complex-coefficient-rescaling/verify.py

Expected:

PASS_JP8_COEFFICIENT_RESCALING incremental_real_saved=10338 blocks=1015 columns=716 orientations=2 controls=11

• Both fixed 27,909-event programs are checked. The complete scalar projections cover all 716 live input columns and arbitrary incoming coefficient work. All 591 distinct JP8 add/scale cases are checked; changed blocks execute literal primitives.
• A separate implementation evaluates 18,925 forward candidate blocks and confirms the restricted optimum. All 6,279 touched-register/work basis cases pass. The 1,211 new magnitude-changing edges are distinguished from 2,023 equal-magnitude edges.
• The full negative census retains 49 nonprofitable transitions (28 ties, 21 losses) and 1,015 incompatible odd-numerator transitions. Backward has no eligible join in this class. No broader circuit-optimality claim.
• Eleven primary controls, four independent semantic controls, and nine additional tests pass, including all 343 signed triples over a seven-value domain, source/receipt corruption, early-ended helpers, relocation and gzip recompression.
• CPython 3.14.7 and 3.11.14 on macOS arm64 pass, including isolated optimized execution. Ruff, local documentation links, staged whitespace and make selected-record-check pass.
• Full make verify, selected cold replay, global admission, Linux CI and Lean were not run for this isolated research addition.

Provenance and coordination

Exact base: 66849e1475c930fd7732b1a0067ffee7e1e95f60. Selected #626 source: 660de893d27bb7abd1d5b01b1640207528a5c851; #622 construction: f69477f4f6d3c7eb1f740b166881bffed4ffa9f6; retained #543: 2065fe4f72a914d907f1046603f5a36f17018535. Full gzip payloads and exact plain-file bytes are pinned; native/event hashes additionally match the preserved #626 SPEC.

OS assignment seed 39, complex-supplier track, selected JP8 by parity. Three candidates and the claim are recorded locally because no pinned coordination issue exists. The final overlap refresh covered 556 open PRs through #737. #722 is the direct methodological predecessor and comparison. #696/#723 and concurrent #734 optimize individual coefficient preparations; this uses ordinary Horner preparation and shares work between macros. No combined saving with those lowerings is asserted. #681, #698 and #719 cover distinct factory, frame and CK rank questions.

Credit to @ellipsi2's #722, Jeremy Kothe/AtomScout and the #622 lineage, Chafik Boukhalfa's #626 package, icekylinx's scalar/emitter work, DreamingOfClouds, Dugongue, eumemic (including #411), Sagar Thakrar, Chris Pagolu and all retained predecessor notices. Prepared for Jeffrey Cox (@namredips) with substantial OpenAI Codex assistance, including one scoped AI reviewer who supplied the independent checker. Apache-2.0; no independent human-review claim.

[Mister-Meeseeks]

100-PR batch review at abb9825202dba6f9bc95c7df8d12d86088fdff1a.

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

Reviewed scope: Dyadic coefficient rescaling and sharing lower JP8 forward primitives; exact rows and prefixes, no automatic composition with other proposals.

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.