Tooling: 1.69× faster production R439 omission atlas via packed lanes with identical certificates

[VictorLiwentao]

Change and scope

At base 0b1cccf86e2f5132ed286a26ef0a681a495740ea, production research/r439-omission-atlas-seed81/verify.py still replayed every single-ADD omission with a fresh 661-bit row map. Sagar Thakrar’s opt-in adapter research/omission-lanes-seed92/ (incorporated #1433) already proved sixteen independent omissions in integer bit lanes with identical complete certificates. This PR wires that equivalent rewrite into the production PR1017 atlas used by R439 omission tooling.

The historical per-omission replay(gates,k) helper is retained for baseline and witness checks. Dependent packages that live-pinned the pairwise-era atlas digest are demoted or refreshed: opt-in omission-lanes-seed92 vendors pairwise baseline.py; r439-three-omission-distance demotes its atlas verify.py / README pins to provenance (it has its own independent replay). Leave-alone tooling through #2412 and deferred ck-dyadic-replay-production / scalar-fused-dyadic-production were not reopened.

No certificate κ pins change.

Tooling improvement; no new kappa. Selected frontier untouched.

Result

On Linux / CPython 3.12.3 with taskset -c 0, five alternating fresh-process pairs over the complete 5,563-mutant atlas gave median wall time 3.630 → 2.143 s (1.694×). Timing is host-descriptive and is not a correctness gate. Every honest receipt matches research/r439-omission-atlas-seed81/CERTIFICATE.json.

Package: research/omission-lanes-production/.
Tip: c36bb5355b97ee82bb2ee0f28b3214fa4583cafb.

Validation

python3.12 -I -B research/omission-lanes-production/verify.py
python3.12 -I -O -B research/omission-lanes-production/verify.py
python3.12 -B scripts/research_package.py --package research/omission-lanes-production
taskset -c 0 python3.12 -I -B research/omission-lanes-production/benchmark.py --rounds 5

Expected:

PASS_OMISSION_LANES_PRODUCTION honest=1 matrices=5563 short_checks=7820 corruption=1 certs=1 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/omission-lanes-production outcome=completed
PASS_IDENTICAL_OMISSION_LANES_PRODUCTION_RECEIPTS

Default verify is read-only. Temp-copy packed-mutant bit-flip foil rejects via audit failure. Package gates under -O still load checkers with compile(..., optimize=0). Full make verify, selected cold replay, Lean, and multiplication-theorem obligations were not rerun for this isolated historical-checker change.

Why the packed lanes are equivalent

Pack a batch of B omissions as width-N lanes inside one Python integer. Lane j occupies bits [jN, (j+1)N). An ADD is an XOR of source into destination; XOR cannot carry across lanes. Suppressing lane j at omitted gate k is exactly the historical if k != omit guard for that lane. See research/omission-lanes-seed92/.

Overlap and credit

Closest priors (different delivery): Sagar Thakrar’s research/omission-lanes-seed92/ / #1433 (opt-in packed lanes; this is the production wiring), CrocSwap/integer-mult-bounds#756. Leave-alone tooling through #2412 and deferred ck-dyadic-replay-production / scalar-fused-dyadic-production were not reopened. No competing open PR on production r439-omission-atlas-seed81/verify.py at refresh.

Credit woklerx and Sagar Thakrar. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.