[codex] Optimize trace2 ingestion side effects

[svarlamov]

Summary

• Move trace2 command side effects out of the latency-sensitive ingest path and schedule them through per-family drains with bounded background concurrency.
• Narrow reflog cursor capture to command-relevant refs instead of scanning every reflog for each mutating trace event.
• Serialize process-local notes writes/merges per repo notes ref to avoid concurrent fast-import/update races.
• Prevent incomplete trace roots from globally blocking unrelated repository families, with unit coverage for that fence behavior.

Root Cause

Trace2 ingestion still did too much synchronous work: broad reflog scans, inline command side effects, and broad family blocking on trace roots that had not identified a repo. Under suite load, an incomplete root could leave later checkpoints waiting behind unrelated trace state.

Validation

• task build
• task lint
• task fmt
• task test CARGO_TEST_ARGS="--lib"
• task test CARGO_TEST_ARGS="--test daemon_mode"
• task test CARGO_TEST_ARGS="--test integration"

Focused regression checks also passed for cold conflict rebase, pull-rebase/autostash, Cursor attribution, partial staging, subdir amend, and large stash note cases.