Distillation

Date: Sep 17, 2026

• 🔴 (09:26) User asserted the embedding pool’s primary worker is always present; it has its own OOM backoff, separate from pool sizing and replacement-worker admission logic.
• 🔴 (09:26) User asserted invalid LORE_EMBED_POOL_SIZE values must resolve to undefined, never NaN; _configuredEmbedPoolSize() in packages/core/src/embedding/pool.ts is a test seam for validating this without spinning up embedding workers.
• 🟡 (09:26) packages/core/src/embedding/local.ts:1067-1096 opportunistically invokes maybeReprobeCap() before each request, applies recovered capacity through effectiveMaxTokens() without respawning the worker, and dynamically clamps maxTokens to current free-memory allowance because native over-allocation can cause an uncatchable SIGKILL.
• 🟡 (09:26) packages/core/src/embedding/local.ts:1071-1095 applies safeLocalTruncate, prefixes inputs with search_document: or search_query: , gives single query-type recall embeddings "high" priority and others "normal" priority, and sends the current effectiveMaxTokens() in each EmbedRequest.
• 🟡 (09:26) packages/core/src/embedding/local.ts:1002-1035 handles synchronous worker respawn failure after WASM fallback by rejecting every pending request with LocalProviderUnavailableError("embedding worker respawn failed after WASM fallback"); resubmitted pending payloads receive a refreshed maxTokens from effectiveMaxTokens().
• 🟡 (09:26) packages/core/src/embedding/local.ts:1145-1186 shutdown behavior: memoizes shutdownPromise, marks the provider closing, rejects in-flight requests with LocalProviderUnavailableError("embedding worker shut down"), awaits initialization, unreferences and shuts down the active worker alongside retired-worker settlement, and throws AggregateError("embedding worker termination was not confirmed") if multiple termination operations fail.
• 🟡 (09:26) Search found 44 references related to embedding pool sizing and memory headroom. Production definitions include PER_WORKER_MEM_BUDGET_BYTES and desiredEmbedPoolSize() in packages/core/src/embedding-cap.ts, liveFreemem() in packages/core/src/embedding/pool.ts, and constrainedMemoryLimit() plus availableMemoryHeadroom() in packages/core/src/embedding/local.ts.
• 🟡 (09:26) packages/core/test/embedding-cap.test.ts verifies PER_WORKER_MEM_BUDGET_BYTES === 1536 * MB; desiredEmbedPoolSize(3 * GB - 1) === 1; desiredEmbedPoolSize(3 * GB) === 2; desiredEmbedPoolSize(64 * GB) === DEFAULT_MAX_EMBED_POOL; explicit maxima are floored (3.9 → 3) and capped by EMBED_POOL_ABS_MAX; and invalid/nonpositive free-memory inputs including Number.NaN, -100, and 0 all produce a minimum pool size of 1.