v1.38.10

🐛 Bug Fixes

• Remove flaky TestReadSnapshotBufferAllocation.
• Size heartbeat digest buffers by set count, not leaf count.
• Read the dimensions bucket once per usage sweep.
• Warm the version map in the background at startup.
• Report the tenant's pre-freeze status on an aborted freeze.
• Unmarshal namespace snapshot outside of lock.
• Size the roaringsetrange pread read buffer from the segment payload.
• Fix flaky teardown in async replication acceptance suites.
• Report one error per object when a shard refuses a batch.
• Move `st.trackDBLoadProgress` inside `st.reloadDBFromSchema`.
• Backport #12079 and #12127 to stable/v1.37 for async-repl.
• Gate tenant actions by namespaces.
• Fix non-compiling test.
• Don't dereference nil target combination in parseNearVec.
• Serve a failed backup's reason from the participant status.
• Keep error type through the objects manager.
• Read only the geo property when resolving a coordinate.
• Surface root-prefilter skips in asyncReplicationStatus.
• Targeted replace scan with newest-wins visibility in lsmkv.
• Keep the parked pread buffer on sequential replace cursor scans.
• Complete updateIndexTenants when one tenant fails.
• Drop per-cycle allocations from the idle compaction path.
• Read targeted-scan ranges straight from the file in lsmkv.
• Parse replace nodes without allocating.
• Fixes bugs in segment group tmp file recovery.
• Include GEO HNSW index in backups.
• Pin replacement replicas to a namespace's home node.
• Treat a non-positive vector cache maxSize as unbounded.
• Hold the posting lock during version map warmup.
• Count occupied slots, and add PreloadIfAbsent to cache.
• Concurrently prefill geo caches using exinsg prefiller.
• Treat an unrecognized task status as in-flight in dtm.
• Bound vector reads by their declared segment.
• Fix Geo property bugs.
• Don't let SIGPIPE break the gh escape-sequence feature check.
• Apply the filter to the BM25 search in Hybrid Aggregation.
• Wait for the local schema before writes that resolve an index by name.
• Fix three crashes and data corruptions in cluster state handling.
• Wait for storage to settle instead of relaxing assertions in usage tests.
• Never publish or load a stale hashtree snapshot in async-repl.
• Don't panic when a batch has objects with no vectorizable properties.
• Reject invalid async replication config before RAFT commit.
• Report every shard failure during index reconcile.
• Fetch Boost.Depth-deep candidates in hybrid sub-searches.
• Skip dropped named vectors when computing cold shard usage.
• Report legacy vector alongside named vectors in usage metrics.
• Make cold tenants report the same data as hot tenants in Usage.