v1.38.9

🐛 Bug Fixes

• Skipped collection tick while previous report is still running in the usage module.
• Avoided reading file sizes in the usage module when not needed.
• Cancelled and drained the cache prefill before teardown in HNSW.
• Fixed 'quantizer not initialized' search error during RQ lazy initialization in flat.
• Avoided holding the DB index lock across a verbose node-status scan.
• Reduced allocations in tenant activity tracking.
• Avoided allocating a mutex on every key-locker call.
• Fixed usage module with named vectors where one is a prefix of another.
• Handled ECONNRESET errors in handlePanics function.
• Handled context cancellation in cursor operations to prevent incorrect nil returns.
• Removed temporary files in compaction/cleanup on failure.
• Switched off bloom filters for collection export.
• Resolved replicas locally and never implicitly activated a tenant in async replication.
• Refreshed the memory monitor while shards are read-only.
• Returned the gRPC query profile when a query matches no objects.
• Used direct map access for shard movement teardown.
• Adapted qa_pr.sh log fetch to gh >= 2.97 escape-sequence guard.
• Stopped worker-pool grow path from leaking goroutines in async replication.
• Restored shard after a failed tenant freeze in async replication.
• Waited for shard readiness after restart across read_repair suite in replication tests.
• Removed the unused shardReindexerV3 scheduler.
• Cleaned up partial HNSW commit log on aborted flat->HNSW upgrade (v1.37).
• Pooled zstd encoders in backups.
• Ensured never to store a partially written chunk in backups.
• Returned explicit bm25 b/k1 values of 0 in schema responses.
• Stopped CheckExpectedStrategy from heap-allocating its variadic slice in lsmkv.