Tooling: exact eligibility masks accelerate historical reclamation

[PeterCarp320]

The historical joint-dual compiler repeatedly visits retired slots whose frames cannot serve the requested region. This opt-in, source-pinned adapter uses exact coordinate-membership masks to remove those visits while preserving the complete eligible order, clearing XORs, paid frames and serialized words.

Tooling improvement; no new κ. Compilation is 4.53–4.87× faster than Chafik Boukhalfa's archived PR175 bucket optimization, and 12.50–14.28× faster than the base compiler, on the two shipped axes. PR175 already solves repeated sorting; this contribution adds geometry filtering and compares against its exact source bytes.

Three cyclic-order rounds on CPython 3.14.6 / macOS arm64, sequential fresh processes:

Axis	Base compilation	PR175 buckets	Coordinate masks
h=23	54.310 s	19.668 s	4.346 s
h=25	93.476 s	31.894 s	6.548 s

The complete measured process, including source admission, serialization, independent dirty-basis replay and transition reconstruction, improves 3.02–3.30× relative to PR175. Raw wall/CPU samples, source and implementation hashes, preliminary probes, and the separate successful public reproduction are retained. These are local historical-verifier measurements.

Validation:

• All 18 primary and six public-reproduction timing processes reproduce both frozen words byte for byte after decompression, and match every compiler, independent replay and transition receipt field.
• Untimed sorted-set audits compare the entire eligible sequence at all 71,519 acquisition requests, including candidates beyond the chosen one.
• 12 focused tests and 16 unchanged historical controls pass. Controls cover stale frame updates/removals, incorrect geometry/rank masks, source drift, disabled assertions, unpaid scatter, missing XORs/transitions and understated width. A deliberately omitted update hook fails real compilation.
• The selected-record/source-inventory check, measured implementation hashes and staged whitespace check pass. Existing sources, manifests, certificates, default commands and upstream/ are unchanged.

Reproduce with standard-library Python 3.11+, from a full checkout and using a fresh output directory:

python3 -B research/reclamation-priority-performance/verify.py --out /tmp/reclamation-review

Expected final marker: PASS_RECLAMATION_PRIORITY. Review map, proof and evidence.

Exact base: 66849e1475c930fd7732b1a0067ffee7e1e95f60. OS-random assignment seed: 85, verification engineering. No sprint coordination issue exists; the claim and scope refinement are recorded locally. Final overlap checks cover 586 open PRs through #767 plus arrivals through #772. Recent tooling #736/#752/#756 addresses different workloads. PR175 at f1081fabbfadaaaaf7292a40d428dbbb865ce920 is the credited, incorporated predecessor and comparator.

The full repository suite, selected cold supplier replay, C++ CRT profiles, assembly and Lean were not rerun for this isolated adapter. Remote CI is separate from local verification. Inherited all-size/compiler/physical/hosting/precision/tape/analytic obligations remain; this is not a formal multiplication theorem.

Prepared for Peter Carpenter with substantial OpenAI Codex assistance. Credit eumemic's compiler and PR154 bottleneck report, Chafik Boukhalfa's ranked reclamation and PR175, Rohan Gupta's PR55 producer, rfu08's PR64 regression work and the retained predecessor lineage. Apache-2.0; full attribution and AI disclosure are in NOTICE.md.

[Mister-Meeseeks]

100-PR batch review at daa9b0e773b53f36d5b4c3f1c7524600e5714deb.

This remains open after scoped triage. Next review step: Complete eligibility audit and h23 reference/optimized replays pass, but the broader command exceeded 180 seconds. Finish h25 and remaining receipt comparisons in a dedicated job; no performance ratio independently certified.

This is a specific continuation point, not a rejection or a request to redo already recorded checks. Completed local evidence is retained separately from outstanding whole-head obligations.

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