v1.37.15

🐛 Bug Fixes

• Route shard access through Index.withShardOrRemote.
• Allocate per-tenant vector cache memory lazily and proportionally to tenant size.
• Unify the loading-shard readiness check and fix FetchObjects.
• Add retry to module base client.
• Fix data race in TestPartialChunkRecovery.
• Pin the nearObject anchor in the unlimited vector search tests.
• Stop truncating UUIDs in id_as_bytes and batch-delete uuid.
• Put actionable error text in msg on the boot and exit paths.
• Reduce HNSW snapshot allocations.
• Require every Bucket to carry a logger.
• Fix disposable_created size label off-by-one.
• Roll back shouldAbort when ctx expires while callback runs.
• Roll back abort when ctx expires while callback runs.
• Skip collection tick while previous report is still running.
• Cancel and drain the cache prefill before teardown.
• Fix 'quantizer not initialized' search error during RQ lazy initialization.
• Don't hold the DB index lock across a verbose node-status scan.
• Reduce allocations in tenant activity tracking.
• Don't allocate a mutex on every key-locker call.
• Parallelize BM25 block term creation across properties.
• Slab-allocate visited sparse-set segments.
• Respect per-query concurrency budget in compressed rescore.
• Fix usage module with named vectors where one is a prefix of another.
• Handle ECONNRESET errors in handlePanics function.
• Handle context cancellation in cursor operations to prevent incorrect nil returns.
• Remove tmp files in compaction/cleanup on failure.
• Resolve replicas locally, never implicitly activate a tenant in async replication.
• Return the gRPC query profile when a query matches no objects.
• Adapt qa_pr.sh log fetch to gh >= 2.97 escape-sequence guard.
• Clean up partial HNSW commit log on aborted flat->HNSW upgrade (v1.37).
• Parallelize rescoring with budget-aware workers.
• Read rescore vectors through one bucket view with pooled buffers.
• Return explicit bm25 b/k1 values of 0 in schema responses.
• Remove TestReadSnapshotBufferAllocation which was flaky.
• Warm the version map in the background at startup.
• Backport #12079 and #12127 to stable/v1.37.
• Targeted replace scan with newest-wins visibility.
• Read targeted-scan ranges straight from the file.
• Fixes bugs segment group tmp file recovery.
• Hold the posting lock during version map warmup.
• Count occupied slots, and add PreloadIfAbsent.
• Bound vector reads by their declared segment.
• Don't let SIGPIPE break the gh escape-sequence feature check.
• Wait for storage to settle instead of relaxing the assertions.
• Don't panic when a batch has objects with no vectorizable properties.
• Add MUVERA specific usage calculations.
• Register disable_dimension_metrics as a runtime override.
• Restore cache reads for uncompressed non-MUVERA late-interaction rescore.
• Write WAL tombstone pair before publishing inverted doc tombstone.
• Fix all linter issues coming form golangci-list v2.13.1.
• Upgrade modules light and modules api runners to ubicloud-standard-4.
• Merge remote shard aggregations after JSON round trip (mean/median/mode panic on multi-node).
• Pin minimum openssl version so cached builds pull CVE fixes.