Tooling: 9.31x faster exact four-flat maximal-mask filtering

[tekuuu]

Change and mathematical scope

The four-flat decoder checker removes contained setup masks by comparing every candidate against every other candidate. This adds an opt-in cardinality-bucket filter: compare a candidate only with already retained, strictly larger maxima, then restore the original mask order. A short induction proves exact maximal-set and ordered-output equivalence.

Tooling improvement. No new pruning rule, decoder exclusion, supplier or kappa. Both measured engines retain PR2975's same option generator. The change targets a different setup bottleneck.

Same-input workload	Reference	Accelerated	Speedup
Maximal-mask filtering, both representatives	104.85 s	11.26 s	9.31x
Complete per-representative setup	107.28 s	13.55 s	7.92x
Full bounded run, shared catalogue/orbits included	156.74 s	61.42 s	2.55x

Complete setup-state bytes agree for all 2,572 active direction records and 2,380,184 maximal masks. All eleven downstream requirement records, the complete output/histogram reporting 5,509 closure leaves, and the known relaxed span-five positive witness also agree. Only compact receipts are committed.

Task and overlap

Refined replay-runtime, OS seed94. Scientific base b3f3ee0392b71b118fd086a6eaf6b4cab54ef455; synchronized submission base 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11. Revision6 guidance was read; this card, selected record and all seven runtime inputs are unchanged.

Exact dependencies: PR2975 1d7b491c9efad451d49dbfd36a3be7443999bf17, retaining DreamingOfClouds' help package 1b25759bff5486293a9083ebbe0097b5f6642e08. Scope is representatives46/94 setup for (a,b,c,d)=(4,6,4,5), the existing twelve representative46 pairs, and the known (6,6,4,0) positive control. Claim. PR2985's residue-closure/input-coverage work is distinct and is not assumed here. No full class-C campaign is launched.

Validation

python3.12 scripts/research_package.py --package research/four-flat-maximal-masks

Expected PASS_MAXIMAL_MASK_CONTROLS, PASS_MAXIMAL_MASK_FULL, and registered-package success. Requires GNU C11/unsigned128 cc and sanitizer-trap support; Python standard library only.

• Registered Python3.12 controls and complete differential replay pass.
• Independent explicit-set oracle: all 65,536 four-point families plus 2,128 wide-mask cases; sanitizer-trap run agrees.
• Six implementation mutants reject; source-byte, adapter-anchor, invalid-input and complete-export controls pass.
• Read-only Git-free optimized Python3.14 controls agree. Publication-byte preflight passes.

Timings are serial local Linux/GCC measurements, not a universal performance guarantee or multiplication-runtime claim. The full comparison shares predecessor catalogue, orbit, feasibility and rank code; those mathematical premises, RRC and supplier/all-size obligations remain unaudited here. Hosted CI and independent human/formal-kernel review are separate.

Credit

DreamingOfClouds for the four-flat engine, original mask construction, request and #2816/#2826/#2862 lineage; PR2975 for the separately retained requirement-enumeration improvement. Tekalign (@tekuuu), with substantial OpenAI Codex assistance in analysis, implementation, checks and writing. Apache-2.0. Frozen predecessors and shared code are unchanged.

Open next questions

The package contains the proof, exact pins, receipts and next questions (10 files, 900 added lines; the filter is 38 lines). A stronger exact orbit reduction or common-flat gate is still needed to make the full campaign affordable; the smallest next experiment is a bounded slice with a proof that the gate preserves every possible span witness.