MLflow tracing with async Stop hook (opt-in) — resolves #9

Summary

Fix for #9 (MLflow tracing silently disabled). Migrating from datasciencemonkey PR #139.

• Opt-in tracing via MLFLOW_CLAUDE_TRACING_ENABLED=true in app.yaml — keeps default behaviour unchanged for existing deployments, but gives users a single env-var to flip.
• Stop hook delegates to mlflow-trace-stop.sh, which backgrounds the handler via nohup timeout 30 … & disown. Returns in <1s so the rest of the Stop hook chain isn't blocked.
• Hook-event JSON via temp file captured synchronously before backgrounding — naive nohup would redirect stdin to /dev/null and the handler would lose the transcript path.
• Hard 30s ceiling on the backgrounded flush prevents a wedged handler from leaking memory/CPU.
• Pins mlflow-skinny and mlflow-tracing to 3.11.1 to match the Apps runtime — version mismatches caused silent import failures.

Why this resolves #9

#9 documents that MLFLOW_CLAUDE_TRACING_ENABLED="false" is hardcoded today; the upstream Stop hook short-circuits and no traces are written despite the README claiming auto-tracing. This PR introduces the env-var override so users can flip it on without modifying source, and adds the async wrapper so enabling it doesn't slow session teardown.

Branch

feat/mlflow-tracing — about to be pushed.

Diff scope

+128 / -32, 3 files. Tests in tests/test_mlflow_tracing.py.

Start with app.yaml, the Stop hook, and mlflow-trace-stop.sh, then run tests/test_mlflow_tracing.py. Verify that opt-in tracing uses the documented environment variable, the hook returns promptly while the handler is bounded, and the pinned MLflow versions match the Apps runtime.