fix(tracing): harden offline buffer for concurrency and high volume

[viniciusdsmello]

Summary

Hardens the offline trace buffer against multi-process / multi-thread use and high volume, where it could silently lose data or grow unbounded past its configured cap. These issues were found while porting the feature to the TypeScript SDK and validating it end-to-end (worker threads + child processes + real HTTP); the same root causes exist here. (Note: the configure() "replace-all" footgun is already fixed on main via the merge-semantics init(), so it is not included.)

Fixes

1. Eviction cap not enforced across processes — store_trace removed only one oldest file per call. With multiple processes sharing the default ~/.openlayer/buffer, concurrent writers each drop one while many add, so the cap is never enforced (reproduced: 4 processes × 800 writes to a cap-200 buffer left 220+ files). Now evicts the full excess in one batch down to max_buffer_size. Also stats each candidate defensively, so a file removed by another writer mid-scan no longer raises and aborts the store (which previously dropped that trace).

2. Filename collisions at high volume — filenames used str(uuid.uuid4())[:8] (32 bits). Across processes the PID disambiguates, but worker threads share a PID, so birthday collisions become likely at tens of thousands of buffered traces → silent overwrite. Now uses the full UUID (122 bits).

3. max_buffer_size=0 coerced to 1000 — max_buffer_size or 1000 treated an explicit 0 as falsy. Now uses an explicit None check and honors 0 (store nothing).

4. clear_buffer aborted on the first failed unlink and could mis-report the count (returned 0 even after deleting some). Now isolates each removal and returns the count actually removed.

5. replay_buffered_traces(max_retries=0) was a silent no-op — range(0) never ran the loop, so traces stayed buffered with a misleading 0/0 result. Now always attempts each trace at least once.

What is NOT changed

• get_buffered_traces already uses sorted(key=...) (one stat per file) and per-file try/except, so no perf/robustness change was needed there (the TS port needed it because JS sort calls the comparator repeatedly).
• Replay still relies on the SDK client's built-in retry/backoff between attempts (unchanged).

Tests

Adds 4 regression tests in tests/test_offline_buffering.py (batch eviction, max_buffer_size=0, resilient clear_buffer, replay(max_retries=0)) — each verified to fail on the prior implementation and pass with the fix. Full tracer test suite (test_offline_buffering, test_tracer_configuration, test_tracing_core) passes (83 tests). ruff check/ruff format clean; mypy introduces no new errors vs. baseline.

🤖 Generated with Claude Code