Review

LGTM. Reviewed the full diff against main (21 files, +154/-109).

What's good

• Efficient CRC threading: The chained CRC flows through the existing read path (codec → segment → WAL → reader → controllers) with zero redundant reads. Callers that don't need the CRC simply discard it with _.
• AppendAndSync callback design: Capturing LastCrc() after appendAsync0 while still holding t.Lock() is correct — no concurrent append can change the CRC before it's bound in the closure. The doSync/runSync batching is left untouched.
• V1 codec: Returning 0 for CRC is the right behavior since V1 has no CRC support.
• Integration test: parseChecksumMetrics is properly parameterized, and both oxia_dataserver_db_checksum and oxia_dataserver_wal_checksum are asserted for non-zero values that change after additional writes.

Notes

• The WAL checksum gauge is only recorded when resp.Checksum != nil (i.e., when the DB checksum feature is enabled). This ties the WAL metric visibility to the DB checksum feature flag, which seems intentional since they serve the same consistency validation purpose.
• All 21 files are consistent in naming (entryCrc) and interface signatures.