Append cross-architecture verification results to replay-scale-fp-perturbation

[gmliao]

Follow-up for PR #47 (experiment/replay-scale-fp-perturbation).

The topic's Part A currently covers arm64-record → arm64-replay only (41 recordings / 70,800 ticks, 0 mismatches). The cross-architecture claim (arm64 → x86_64) still rests on the 2026-02 evidence, which was produced by an older code version whose records no longer replay on current game logic — so it should be re-established on the current batch.

Steps

1. Copy reeval-batch-arm64.tar.gz (prepared on the recording machine, ~2 MB, 41 records) to the x86_64 (WSL/AMD) machine and extract: mkdir -p /tmp/reeval-batch && tar xzf reeval-batch-arm64.tar.gz -C /tmp/reeval-batch
2. On branch experiment/replay-scale-fp-perturbation (or main after the PR merges): bash deep-research/replay-scale-fp-perturbation/crossarch-verify.sh
3. Expected: cross-arch verify (x86_64): pass=41 fail=0
4. Append the results to the topic (new crossarch-* artifact JSONs + README table row + Caveats update), keeping regenerate.py --check green.

Notes

• Records are deterministic: they can also be regenerated on the arm64 machine via ReevaluationRunner --record --seed-id <N> ... (seeds 1–30, 101–103, 111–113, 121–123 with the matching --players/--move-every, 201–202 with --ticks 12000).
• Related follow-ups tracked separately: core replay sequence-counter fix, and a framework-level deterministic collection type (StateTable) — see the design note in Notes/plans/2026-08-31-replay-scale-fp-perturbation-experiment-design.md.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Sm3UKGwyUw6RyELpBMFd3H