Distillation

Date: Sep 17, 2026

• 🔴 (09:49) User stated _configuredEmbedPoolSize() must handle invalid configuration/environment values by returning undefined, never NaN, and must do so without spinning up a worker.
• 🔴 (09:49) User stated the primary embedding worker is always present; its own OOM backoff, rather than pool sizing, governs its recovery.
• 🔴 (09:49) User stated telemetry must never break the embedding path.
• 🔴 (09:49) User stated the last self-heal re-probe timestamp uses epoch milliseconds, with 0 meaning never and causing it to be primed on the first tick.
• 🔴 (09:49) User stated worker entrypoint resolution differs by module format: CJS uses __filename, which is always defined, while ESM uses import.meta.url.
• 🟡 (09:49) Embedding-pool code defines MAX_NORMAL_QUEUED_EMBED_OPERATIONS = 240.
• 🟡 (09:49) The embedding worker stream processes 256 items per round trip to maximize throughput without blocking the event loop.
• 🟡 (09:49) Worker initialization passes a cgroup-CPU-aware native ONNX Runtime intra-op thread cap, computed on the main thread, through WorkerInitData.
• 🟡 (09:49) Worker entrypoint resolution supports both a SEA-extracted worker asset and a parent-level TypeScript worker in the source tree, and matches the sibling worker file extension to the current bundle.
• 🟡 (09:49) WASM-fatal or OOM worker errors trigger rejection through the OOM-handling path; recovery respawns a fresh worker and re-submits work.
• 🟡 (09:49) Transient initialization failures retry a fresh worker after a cooldown rather than permanently disabling the local embedding provider.