Tooling: 3.15× faster production Clifford zero_state via gate tables with identical receipts

[VictorLiwentao]

Change and scope

At base 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, production multicold clifford_exact.py still built the exact zero state by expanding the Clifford circuit. The opt-in adapter research/clifford-state-performance/ (lineage of #662) already proved gate-table permutation/phase evaluation with identical cold617 helper-interface receipts. This PR wires that equivalent zero_state into the production multicold evaluator used by scripts/verify_cold617_interface.py.

Path: research/complex-multicold/vendor/pr558/vendor/pr552/complex-global/vendor531/vendor/clifford/clifford_exact.py.

No certificate κ pins change. Same-tip provenance hygiene:

• research/clifford-state-performance/ vendors pairwise-era bytes as baseline.py and demotes the live pairwise pin.
• research/clifford-reciprocal-basis-obstruction/ demotes clifford_exact to provenance and drops it from live VERIFY.json inputs.
• research/cold-wrapper-amplitude-audit/ moves the pairwise clifford_exact digest to provenance_sha256 (that audit does not import the module).

Tooling improvement; no new kappa. Selected frontier untouched.

Result

On Linux / CPython 3.12.3 with pinned CPU, five alternating fresh-process pairs gave median helper-interface wall time 1.202 → 0.382 s (3.15×). Timing is host-descriptive and is not a correctness gate. Helper-interface receipts match the baseline evaluator (states=100, instances=2).

Package: research/clifford-state-production/.
Tip: 68a7161a610f175dc8e7b22f95acd033191450d9.

Validation

python3.12 -I -B research/clifford-state-production/verify.py
python3.12 -I -O -B research/clifford-state-production/verify.py
python3.12 -B scripts/research_package.py --package research/clifford-state-production
python3.12 -B scripts/research_package.py --package research/clifford-reciprocal-basis-obstruction
python3.12 -B research/cold-wrapper-amplitude-audit/verify.py
python3.12 -B -m unittest tests.test_clifford_state_production

Expected:

PASS_CLIFFORD_STATE_PRODUCTION states=100 instances=2 corruption=subprocess_failure certs=1 no_new_kappa
PASS_REGISTERED_RESEARCH_PACKAGE research/clifford-state-production outcome=completed
PASS_REGISTERED_RESEARCH_PACKAGE research/clifford-reciprocal-basis-obstruction outcome=completed
PASS_COLD_WRAPPER_AMPLITUDE_AUDIT …

Helper-interface receipt/benchmark runs briefly swap the live production clifford_exact.py under try/finally restore, because scripts/verify_cold617_interface.py forces MULTICOLD_SOURCE onto the repository multicold tree (temporary overlays are ignored). The working tree is restored after every run. Permutation-foil rejects via subprocess failure. Package gates under -O still launch assertion-enabled children. Full make verify, selected cold replay, Lean, and multiplication-theorem obligations were not rerun for this isolated evaluator change.

Overlap and credit

Closest priors (different delivery): research/clifford-state-performance/ / #662 (opt-in gate tables; this is the production wiring), CrocSwap/integer-mult-bounds#756, CrocSwap/integer-mult-bounds#2257 / #2263 / #2276 / #2320 (same production-promotion + provenance pattern). No competing open PR on this production clifford_exact.py at refresh.

Credit Charles Bethin and the multicold Clifford lineage. Prepared by Wentao Li (VictorLiwentao, @Is__Victor) with substantial AI assistance, under Apache-2.0.

[Mister-Meeseeks]

Maintainer review of 68a7161a610f175dc8e7b22f95acd033191450d9: the gate-table rewrite has a plausible exact recurrence, and the recorded normal/optimized package replays pass. Before production integration, please fix the runtime source-binding regression.

clifford-reciprocal-basis-obstruction/dense.py:native_module() still imports V.ROOT / V.KERNEL, but this PR removes that actual runtime dependency from INPUTS, VERIFY.json and CERTIFICATE.json, marking it as provenance. It also changes the former kernel-drift negative control to mutate selector-batching/README.md instead. I reproduced that appending bytes to the executed Clifford oracle is now accepted by check_sources() with an unchanged reported source inventory.

A consumed oracle is an input, not historical provenance. Please preserve a pinned original runtime baseline for the historical replay, or explicitly migrate the replay to a pinned optimized oracle with complete regenerated evidence and the original drift-control coverage. Keep immutable evidence intact and audit the same issue in cold-wrapper-amplitude-audit, whose executed-source pin is also removed. Include a control proving kernel source drift and missing kernel inventory are rejected; do not substitute a README mutation for that coverage.

This is a reproducibility/integration hold, not a rejection of the speedup. Exact evidence is retained in the efficient1300 audit as combined-2329-source-control.py/.json, to be published with this checkpoint.