DONE: native checker compilation reuse with complete execution

[bitplane]

DONE / 2026-10-15 08:18 UTC / claim released.

Published #2414, exact head 160c6b06954339f97d53529557247c22ff27c47d.

A warm private ccache reduces the measured local build-and-replay median for this one unchanged nine-stage checker/two-word fixture from 2.753 s to 0.240 s (11.46x); first fill is 2.942 s. Every sample freshly links and runs both complete words. All native receipt fields agree after only elapsed-time removal and content-bound input-path normalization. No production entry point, mathematical checker, selected certificate or kappa changes.

Executed 13 builds and 24 native runs, retaining 384 omitted-bridge controls across successful runs. Real source/header/option/profile invalidation, restored-input hits, an invalid word after a hit and corrupted-cache recovery all pass. Registered Python 3.12 and nine package control groups pass. Full Python 3.14 native replay from a Git-free read-only 18-file execution snapshot gives identical normalized native results and leaves the snapshot unchanged (2.761 s / 0.241 s medians).

The checker belongs to the DaysSky / Dylan Crow / cohort lineage; raw fixtures come from our PR2373 and Sagar Thakrar's PR2115 construction. Ccache is the existing external implementation, not a new algorithm. This is one compilation-boundary experiment, not the full selected replay or a new physical supplier. Exact commands, environment/receipt hashes and remaining limits are in the package. Current hosted CI snapshot: {'QUEUED': 62}; success is not claimed.

Open handoff: integrate this opt-in compile/object-cache/fresh-link boundary into a named native harness while retaining its exact pins, legality and 5/9/11-stage executions; then measure the whole pipeline. Substantial OpenAI Codex assistance disclosed; no subagents or independent human review.

---

Original claim retained:

CLAIM / 2026-10-15 07:52 UTC / priority revision 4.

Question: can repeated compilation of the unchanged native nine-stage BIT column checker be reused while every mathematical check still executes? Pilot: one GCC build took 2.82 s; the complete native nine-stage checks of the two pinned PR2373 words took 0.0045 s and 0.0021 s. This is compilation reuse, not cached mathematical success.

Base: 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2. One OS draw 2f939a000f93b5f7feca53d415dcfc3b selected verifier-engineering / replay-runtime / source-bound reuse. The required selected baseline passed. Exact source: merged research/e8-bit-carrier-skeleton/code/bit/checkers/columns_k.cpp; native fixtures from our PR2373 at b52f87f5355ad9f02ca4ac4395117655ac1b19c7, retaining its source lineage and all 1,327 columns, nine stages and 18 omitted-bridge controls per word.

Method/boundary: opt-in ccache preprocessor-mode object reuse, fresh linking and native execution every time. Authenticate source bytes and fixtures; explicit compiler/toolchain profile; test header/source/compile-option/profile invalidation and wrong mathematical input after a cache hit. Compare complete before/after native receipts except their declared timing field, with identical fixtures. Measure uncached, cache-fill and warm-cache time separately. No production entry-point or immutable input changes, new supplier, whole selected cold replay or kappa claim.

Expected artifact: a scoped executable package with benchmark commands, exact provenance, retained checks and actual invalidation controls; otherwise a bounded negative result. The fresh complete overlap snapshot has 611 open PRs and 1,137 Discussion records, without a matching compiler-cache investigation in the targeted terms. Metadata coverage is not proof of novelty.

Evidence is LOCAL-ONLY until publication. Claim expires 2026-10-15 09:52 UTC.

Gareth Davidson (@bitplane), with substantial OpenAI Codex assistance. No subagents or independent human review.