fix: stabilize concurrent trace snapshots

[Iams4kura]

What changed

• Guard every in-process TraceStore accessor with the existing reentrant lock.
• Capture record references as an immutable snapshot under the lock, then deep-copy outside the critical section.
• Add a deterministic reader/writer regression test that freezes snapshot copying while another thread appends.

Why

TraceStore.append() updates the in-memory record index while holding _thread_lock, but records() iterated the same dictionary without that lock. A concurrent append could resize the dictionary mid-iteration and raise RuntimeError: dictionary changed size during iteration.

Impact

Readers now observe one stable same-instance snapshot. Large deep copies do not block appends after snapshot references are captured. On-disk JSONL format, cross-process flock behavior, and resume semantics are unchanged.

Validation

• Regression reproduced before the fix.
• TraceStore suite: 8 passed.
• Concurrent regression: 20/20 local repetitions; 30/30 independent review repetitions.
• make ci: Ruff and strict mypy; 245 tests passed with 11 optional Qwen tests deselected; 64 smoke evaluation units; 20 deterministic semantic canary repetitions; reproducible sdist/wheel and isolated installed-wheel contract.
• Independent diff review found no blocking issues.