Tooling: 2.87× faster ternary verify_invocation_frames with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, the historical ternary-side small controls (scripts/audit_ternary_side.py on h=8,9) spend almost all wall time in scripts/experiments/ternary_side.py:verify_invocation_frames. Profiling shows repeated Fraction matrix products/inverses while building Gram projections, and repeated rank(sub(F, old)) work on recurring physical edges.

This PR keeps the default checker path and:

1. clears denominators once per Gram factor for products and clears the pivot row before Gauss–Jordan scaling for inverses (same rational values as the historical entrywise Fraction arithmetic);
2. memos exact ordered frame transitions for one verify_invocation_frames call by identity-then-value interning;
3. caches projections on immutable row tuples within the call.

Every support/topology require, both orientations, and every orientation evidence field stay identical. Only the scripts/experiments/ternary_side.py source pin changes inside certificates/ternary-side.json. The h=29 DAG construction and assembly invoice are not replayed. Open CrocSwap/integer-mult-bounds#1922's audit_joint_frames.py edit is left independent; this answers that PR's open ternary inverse question on the ternary frame stream only.

Tooling improvement; no new kappa. Selected frontier untouched.

Result

On Linux / CPython 3.12.3 (Linux-6.12.94+-x86_64-with-glibc2.39), five alternating fresh-process pairs timing build(..., retain_graph=True) plus verify_invocation_frames for h=8 and h=9 gave median wall time 13.578 → 4.733 seconds (2.87×). Timing is host-descriptive and is not a correctness gate. Every orientation receipt matches the baseline checker fieldwise.

Package: research/ternary-frame-cleared-cache/.

Validation

python3.12 -I -B research/ternary-frame-cleared-cache/verify.py
python3.12 -I -O -B research/ternary-frame-cleared-cache/verify.py
python3.12 -B scripts/research_package.py --package research/ternary-frame-cleared-cache
python3.12 -I -B research/ternary-frame-cleared-cache/benchmark.py --rounds 5

Expected:

PASS_TERNARY_FRAME_CLEARED_CACHE frame_cases=2 product_cases=160 certs=1 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/ternary-frame-cleared-cache outcome=completed
PASS_IDENTICAL_TERNARY_FRAME_RECEIPTS

Default verify is read-only. Cleared-product scale-omission and poisoned transition-memo foils reject. Full make verify, selected cold replay, Lean, and multiplication-theorem obligations were not rerun for this isolated historical-checker change.

Overlap and credit

Closest priors (different objects or delivery): CrocSwap/integer-mult-bounds#756, CrocSwap/integer-mult-bounds#1217 (opt-in ternary frame-pair memo; credited, not re-opened), CrocSwap/integer-mult-bounds#1922 (cleared product + open ternary inverse question; audit_joint_frames.py left independent), CrocSwap/integer-mult-bounds#1793, CrocSwap/integer-mult-bounds#1861, CrocSwap/integer-mult-bounds#1984, CrocSwap/integer-mult-bounds#1994. No competing open PR on the default verify_invocation_frames path at refresh.

Douglas R. Colkitt authored the ternary checker; Zhihao Chen’s #7 supplied the five-subset motif. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.