Distillation

Date: Sep 16, 2026

• 🔴 (23:54) User stated packages/core/test/embedding-pool.test.ts “never touches” the per-worker OOM/self-heal lifecycle; that lifecycle is covered separately in embedding-oom-recovery.test.ts.
• 🟡 (23:54) packages/core/test/embedding-pool.test.ts is a 1,650-line Vitest suite for EmbeddingPool cross-worker dispatch issue #999, covering least-busy routing, lazy and memory-gated growth, the broken-provider latch, and pool shutdown.
• 🟡 (23:54) packages/core/test/embedding-pool.test.ts defines a controllable FakeWorker extends EventEmitter using the _setTestWorkerFactory seam. It records embedIds and embedPostCount, tracks gotShutdown and terminated, supports throwOnNextEmbed, and provides completeAll(), completeNext(), startNext(), initError(), crash(), and exit() controls.
• 🟡 (23:54) FakeWorker.postMessage() throws Error("worker died before postMessage") when throwOnNextEmbed is set; embed results contain vectors: [new Float32Array([1, 0, 0])]; shutdown normally emits exit with code 0; terminate() marks the worker terminated and resolves to 0.
• 🟡 (23:54) Test helpers in packages/core/test/embedding-pool.test.ts include installFakeWorkers(), flush() with two sets of 8 microtask flushes around setTimeout(..., 0), flushFakeTimers() using vi.advanceTimersByTimeAsync(0), settle<T>() for immediately handled promise outcomes, and warmPool() using embed(["bootstrap"], "document").
• 🟡 (23:54) The suite’s beforeEach() forces the local provider by removing VOYAGE_API_KEY, OPENAI_API_KEY, and LORE_EMBED_POOL_SIZE; neutralizes real CI cgroup limits with _setConstrainedMemoryForTest(0); resets the local-provider probe; and saves/clears the provider.
• 🟡 (23:54) The suite’s afterEach() restores worker factories, pool sizing, watchdogs, cooldowns, memory test overrides, provider state, VOYAGE_API_KEY, OPENAI_API_KEY, LORE_EMBED_POOL_SIZE, and NODE_ENV; defensively calls _setRecallEmbedsInFlightForTest(0); resets shutdown tracking; and restores real timers.
• 🟡 (23:54) Test serializes cold bootstrap before enabling parallel dispatch configures pool size 2 and 64 * GB free memory. Two initial concurrent requests (alpha, beta) use only one worker and dispatch one native batch until a real result proves model readiness, preventing sibling initializers from racing the shared model cache; after bootstrap, concurrent gamma and delta requests grow the pool to 2 workers and dispatch one request to each.
• 🟡 (23:54) Test does not spawn a second worker for sequential (non-concurrent) embeds configures pool size 2 and 64 * GB free memory, then verifies sequential alpha and beta requests reuse worker 0 and do not load a second model.
• 🟡 (23:54) Test removes a cancelled queued request without restarting its worker configures pool size 1, uses 64 * GB free memory, warms the pool, queues blocked and cancelled requests, aborts cancelled through an AbortController, and verifies only one request reached the existing worker before cancellation.