perf: repair comparator evidence contracts

[kim-em]

Closes #9804.

Repairs the three rolled-back Phase 4 comparator evidence contracts:

• reruns HexBareiss with a warmed FLINT process, an in-harness steady-state overhead case, 12/12 eligible agreeing pairs, and the normative fraction-free versus multimodular-CRT rationale;
• warms both arms of all 44 HexPoly/FLINT pairs, hoists fixture preparation outside timed closures, registers synchronous overhead, and publishes a clean 89-registration/445-repeat export;
• corrects HexGF2's operation-specific NTL contract, keeps seven addition pairs as correctness/protocol anchors, warms all 60 pairs, registers synchronous overhead, and publishes a clean 121-registration/553-repeat export;
• removes the former over-ceiling GF2 GCD n = 1024 and n = 1536 rungs, leaving all 53 performance pairs eligible and below ten seconds;
• extends persistent-process warmup lint coverage to NTL and restores all three libraries independently to Phase 4 with empty live Concerns.

Every artifact records clean source commit f4f013c. All registrations have stable repeat hashes, all pairs have observed-hash agreement, and steady-state overhead-adjusted ratios are reported where the 5% rule requires them.

A fresh Claude Opus second opinion identified fixture preparation, traceability of overhead measurements, ceiling handling, NTL lint coverage, and serialization-sensitive narrative as real issues; all are addressed here. Its proposed payload-dependent overhead definition was not adopted because the binding Process call contract defines the trivial-input round-trip floor, and its objection to NTL half-GCD was rejected after checking the NTL 11.6 source and measured linked build.

Verification:

• lake build (10,500 jobs, after merging current origin/main)
• lake exe hexpoly_bench verify (104/104)
• lake exe hexbareiss_bench verify (28/28)
• lake exe hexgf2_bench verify (144/144)
• python3 scripts/check_phase4.py
• python3 scripts/check_dag.py
• python3 scripts/ci/check_persistent_flint_warmup.py (369 registrations)
• python3 -m unittest scripts/ci/test_check_persistent_flint_warmup.py (13 tests)
• artifact integrity/pairwise-hash/ten-second-ceiling audit
• git diff --check