- Compat tests assert the exported SCHEMA_VERSION instead of a stale
hardcoded 5; the perf-sample change bumped agent.db to v6.
- Backfill cap fixture inserts its 300 rows in one transaction; per-row
implicit transactions fsynced 300 times and timed out on CI runners.
- Verified throughput (TPS) and latency (TTFT) aggregation logic for multiple samples.
- Confirmed data sanitization by dropping invalid samples like NaN, zero duration, or missing values.
- Ensured backfilling correctly filters stale, errored, and empty turns from the stats database.
- Validated that backfill operations respect the maximum sample limit per model.
- Validated the async deferral behavior in recording perf samples.