v0.10.7

Added

• sdl.context defaults to hybrid retrieval end-to-end: hybrid seeding (entitySearch
  via FTS + vector + RRF) now runs alongside path-inference, not instead of it.
  Path-inferred refs are preserved first; hybrid adds semantically related refs
  the heuristic misses. TaskOptions.semantic (default true) and
  TaskOptions.includeRetrievalEvidence (default true) toggle the behavior.
• Retrieval evidence surfaced on sdl.context responses: ContextSeedResult
  gained an evidence?: RetrievalEvidence field, captured from Stage 1
  entitySearch and propagated into AgentContextPayloadSchema.retrievalEvidence.
  The Zod schema now accepts sources, candidateCountPerSource,
  topRanksPerSource, fusionLatencyMs, fallbackReason, ftsAvailable, and
  vectorAvailable — so callers can see which lanes contributed to their
  results.
• Entity-search telemetry: entitySearch emits
  logger.info("Entity search", { eventType: "entity_search", ... }) at every
  return path (fallback-to-legacy, db-connection-fail, main success) with
  latency, per-source candidate counts, and ftsAvailable/vectorAvailable.
  This makes hybrid contribution visible for sdl.context calls (previously
  only sdl.symbol.search emitted hybrid telemetry).
• Multi-model embedding pass: new semantic.additionalModels: string[]
  config option. The indexer now loops over [model, ...additionalModels]
  calling refreshSymbolEmbeddings per model, so both jina-embeddings-v2-base-code
  and nomic-embed-text-v1.5 vectors can be populated from a single index run.
  Hybrid fusion can then RRF-merge both vector lanes alongside FTS.

Changed

• ONNX embedding models unbundled: jina-embeddings-v2-base-code (~162MB) and
  nomic-embed-text-v1.5 (~137MB) are no longer shipped inside the npm tarball.
  Published package size dropped from ~215 MB to ~1.6 MB. Both models are now
  fetched on npm install by scripts/postinstall-models.mjs into the platform
  cache directory (%LOCALAPPDATA%/sdl-mcp/models/ on Windows,
  ~/.cache/sdl-mcp/models/ elsewhere). Set SDL_MCP_SKIP_MODEL_DOWNLOAD=1 to
  skip the postinstall download.
• Model registry fallback URLs: ModelInfo.fallbackDownloadUrls points at
  the project's GitHub Releases mirror (tag models-v1) and is used
  automatically when the primary HuggingFace URL fails at postinstall time or
  on lazy runtime download. Both bundled models now have primary HuggingFace
  URLs populated (jina previously had none).

Fixed

• Hybrid retrieval FTS and vector lanes silently returned zero: The
  QUERY_FTS_INDEX and QUERY_VECTOR_INDEX calls in src/retrieval/orchestrator.ts
  were malformed — missing RETURN clause (rejected by the LadybugDB binder),
  passing the index name as a parameter (must be a literal), and passing topK
  positionally (must be K := <int>). Every invocation threw and the error was
  swallowed at logger.debug level, making both lanes invisible contributors
  with zero candidates. Fixed in four call sites (two in hybridSearch, two in
  entitySearch). The fts.conjunctive config flag is now threaded through to
  the FTS call. Index and table names are validated against a strict identifier
  whitelist before interpolation. Swallow-catch logging bumped from debug →
  warn so future binder errors surface at the default log level.

Breaking Changes

• API Consolidation: Reduced tool surface from 37 to 34 actions
  • Removed sdl.agent.context: Use sdl.context (Code Mode) for task-shaped context retrieval
  • Removed sdl.context.summary: Use sdl.context which provides structured cards + skeletons
  • Merged sdl.symbol.getCards into sdl.symbol.getCard:
    • Single symbol: { symbolId: "..." } (unchanged)
    • Batch: { symbolIds: ["...", "..."] } (was separate tool)
    • Response shape unchanged for each mode

Performance

• Deferred fresh-DB index bootstrap: Fresh databases now create only base schema (node/edge tables)
  during initialization, deferring 19 secondary indexes and retrieval indexes until after the first
  successful full index completes. This removes index maintenance overhead from the ingestion-critical
  path. Existing DB initialization is unchanged.
• Write connection pool: Replaced single write connection with a pool of 4 warm connections using
  two-layer concurrency control (connection acquisition + serialized execution). Enables instant
  failover and hides connection acquisition latency from the batch drain loop.
• Background batch drain: Batch persistence accumulator now uses an async background writer with
  auto-enqueue at threshold (200 rows). Parsing threads are never blocked by DB writes.
• TS engine batch persistence: TypeScript pass1 indexing now uses the same BatchPersistAccumulator
  as the Rust engine, eliminating per-file transaction overhead and enabling write pipelining.

Fixed

• Repository scanning descriptor cleanup: Replaced async node:fs/promises.glob()
  traversal in the indexer with an explicit directory walker that closes every
  opened directory handle deterministically. The same walker now backs Go,
  Java/Kotlin, and C# import-resolution fallback scans to avoid leaving identical
  traversal leak candidates in place.

Commits since v0.10.6

• chore(release): v0.10.7 (0e75955)
• ci fix (ecce466)
• fix(context): surface retrievalEvidence in broad-mode responses + docs (554f176)
• feat(context): default sdl.context to hybrid retrieval end-to-end (5fec465)
• test(ci): align model/package expectations and release tag handling (aca68e9)
• chore(install): unbundle ONNX models and add postinstall downloader (d433a2a)
• fix(retrieval): repair malformed FTS and vector queries in hybrid search (aa748b4)
• Merge branch 'main' of https://github.com/GlitterKill/sdl-mcp (bec9994)
• doc update (8455b6b)
• Remove infographic from iris-gate-ladder.md (f854b1f)
• docs: update mermaid diagram themes and styling across documentation (71c32e1)
• fix(install): add Node.js 24+ startup check and document prerequisites (e5b6447)
• ci fix (89ebaab)