Distillation

Date: Sep 16, 2026

• 🟡 (13:00) Inspection of /home/byk/Code/opencode-lore-embedding-runtime/packages/core/src/embedding-worker.ts lines 110–314 showed isCorruptModelError(msg: string): boolean excludes network/auth/HTTP errors, recognizes protobuf/model parsing, deserialization, and corruption signatures, and must remain synchronized with embedding-worker-types.ts.
• 🟡 (13:00) isNativeOrtLoadError(msg: string): boolean first requires a native-runtime signature such as onnxruntime_binding.node, an ONNX Runtime shared library, or OrtGetApiBase, then recognizes loader failures including ERR_DLOPEN_FAILED, invalid ELF/Mach-O/Win32 binaries, truncated files, missing shared objects/images/modules, undefined or missing symbols, and dlopen.
• 🟡 (13:00) ensurePipeline() in /home/byk/Code/opencode-lore-embedding-runtime/packages/core/src/embedding-worker.ts attempts loadPipeline() lazily. For non-vendored models using the native binding, a native ONNX loader error posts { type: "init-needs-wasm", error: msg }, sets wasmRespawnRequested = true, and returns without posting init-error.
• 🟡 (13:00) Corrupt-model recovery in ensurePipeline() checks cachedModelLooksIntact() before destructive deletion. If files are intact and usedNativeBinding is true, it treats parsing failure as native-runtime incompatibility, requests a fresh WASM worker via init-needs-wasm, and avoids init-error; if already on WASM, it retries loadPipeline() once without purging. If files are not intact, it calls purgeCachedModel() before a redownload retry. Vendored models skip purge/redownload because their model is bundled and read-only.
• 🟡 (13:00) Recoverable native-load, intact-model parse, and purge/redownload diagnostics are gated by stderrSilenced and emitted with console.debug rather than console.warn; the stated escalation rule reserves warnings for actionable/Sentry-warning-worthy events and avoids corrupting host TUI output.
• 🟡 (13:00) Assistant stated it was reconciling the implementation against three must-fix admission findings and newer native-loader regressions, would change only gaps that still reproduce, and would preserve each fix with a targeted mutation proof.