Tooling: cache exact R420 geometry for 6.47x faster mutation replay

[mtpsl]

This verification-tooling improvement accelerates PR543's unchanged mutation suite from 8.239 s to 1.273 s (6.47×) in one same-input local A/B. All diagnostics match byte for byte, and the complete mutation receipts are equal. No construction, exponent or selected-frontier change.

The opt-in adapter caches only a successful exact E8 geometry audit, keyed by the full serialized inputs to the pinned constructor and label checker. Every word still runs the unchanged scalar/frame/dirty-restoration and interface checks. Returned objects cannot mutate the cache; changed geometry is revalidated; failed geometry is not retained. The cache is bounded, in memory, and refuses predecessor source drift.

Only research/r420-geometry-cache/ is added. Existing source packages, manifests, selected certificates and upstream/ are unchanged. This is independent of our earlier circuit PR691.

Reproduce

python3 -I -B research/r420-geometry-cache/verify.py

Expected: PASS_R420_GEOMETRY_CACHE original_controls=10 supplemental_checks=20, followed by fresh timings. Python 3.10+, standard library only.

The command runs the original and cached implementations in separate cold processes on the exact same original word and mutation harness. It preserves the original nine negative controls and positive control, including their rejection reasons. Twenty additional checks cover every geometry key field, eight changed per-word fields, valid reflected E8 geometry, failed-cache admission, input/output aliasing, eviction and source-pin refusal. Exact derived geometry and full successful receipts are compared too.

Local evidence: macOS arm64 / Python 3.12.3. One original-then-cached pair, OS caches not flushed; no timing threshold and no full-pipeline speedup claim. Complete receipts, diagnostics and timings are committed under evidence/. make selected-record-check, staged whitespace validation and the direct-CLI guard pass. The final CLI-only guard was checked separately; the imported implementation measured in the paired run is unchanged.

This scoped optional adapter did not run the full repository suite, selected cold replay, hosted invoices/moments, Lean or remote CI. All inherited all-size/compiler/physical/hosting/precision/tape/analytic assumptions remain. The cache argument is scoped to the pinned checker and ordinary finite JSON words with sequential in-process calls.

Provenance and overlap

Base main: 66849e1475c930fd7732b1a0067ffee7e1e95f60, unchanged at submission refresh. PR543 supplier: 2065fe4f72a914d907f1046603f5a36f17018535. All three execution inputs have SHA-256 and Git-blob pins.

OS seed 91, verification-engineering track. Three candidate tasks and the narrow local claim are recorded. No pinned coordination issue exists. Refresh through #704 found no overlapping PR543 geometry-cache contribution. PR452 addresses earlier PR358 prefix/arithmetic reuse, PR662 complex Clifford helpers, and PR667/672 interval arithmetic; none is imported or reviewed here.

Credit: eumemic supplied the original PR543 exact checker, mutation coverage and R420 word. The README preserves the broader E8, dirty-interface and supplier lineage. Prepared for @mtpsl with substantial OpenAI Codex assistance, Apache-2.0. No independent human review claimed.