Distillation

Date: September 16, 2026

• 🔴 (13:25) User stated the WASM heap never shrinks: after the first oversized allocation grows the heap, every smaller retry in the same worker also fails; real recovery requires replacing or respawning the worker.
• 🔴 (13:25) User stated native ONNX Runtime load-error classification must be narrowly scoped so an unrelated resolution error, such as a missing model file, is never misclassified.
• 🔴 (13:25) User stated formatted-input objects and other later arguments in model errors never match the relevant classifier patterns; classification must inspect only the leading string.
• 🔴 (13:25) User stated indiscriminately treating model parse failures as corruption can cause a purge→redownload→fail→purge loop that never converges and hammers the Hugging Face Hub.
• 🔴 (13:25) User stated the SEA path uses native ONNX Runtime via __LORE_ORT_BINDING_PATH__ and never sets WorkerInitData.forceWasm; forceWasm is set by the main thread only when respawning after a native worker posts init-needs-wasm.
• 🔴 (13:25) User documented native-to-WASM fallback behavior across packages/core/src/embedding-worker.ts, packages/core/src/embedding-worker-types.ts, and packages/core/src/embedding/local.ts: forceWasm defaults to false, a native worker can post { type: "init-needs-wasm", error: msg }, the main thread sets this.forceWasm = true, and respawnForWasm() starts a fresh worker/module graph that skips native ONNX Runtime.
• 🔴 (13:25) User stated embedding-worker-types.ts is imported by both worker and main-thread sides and must remain free of runtime dependencies; error predicates are canonical there but inlined in packages/core/src/embedding-worker.ts to keep the worker self-contained.
• 🔴 (13:25) User documented that the main thread must reject and remove the pending request before respawnForWasm() re-submits work, preventing pending requests from being lost or duplicated during the native-to-WASM transition.
• 🔴 (13:25) User stated corruption healing should purge the cached model and retry download only once, only when the model is not vendored and on-disk validation indicates genuine corruption; an intact model with a native-binding parse failure should be retried without purging or should trigger WASM fallback.
• 🔴 (13:25) User documented ONNX file validation: a candidate file is considered structurally plausible only when sizeBytes >= MIN_ONNX_FILE_BYTES and its first byte is 0x08, the leading ONNX/protobuf header byte; the expected intact model is approximately 137 MB.
• 🔴 (13:25) User stated native ONNX Runtime inference threads are selected on the main thread via nativeIntraOpThreads() and passed to the raw-TypeScript worker as WorkerInitData.intraOpThreads.
• 🔴 (13:25) User documented embedding-worker queue behavior: high-priority work is inserted after the last existing high-priority item, preserving FIFO order within the high-priority class; queued requests are processed one at a time because ONNX inference is synchronous.