v0.1.8: Await embedding warm-up at setup
Await embedding warm-up at setup
Fixes the 0.1.7 regression where 'void embedQuery(warmup)' yielded to the microtask queue and lost a race with @opentelemetry/instrumentation-http installing its fetch wrap. The wrapped fetch can't complete the cold-cache HF Hub download — every fresh container failed its first cyanheads_search with 'Unable to get model file path or buffer'.
Fixed
CatalogService.initialize()now awaits the warm-upembedQuery('warmup', …), keepingpipeline()inside the pre-OTEL setup window where the model fetch succeeds.
Changed
- Server startup blocks on model warm-up — same profile as 0.1.5 (~2 s warm-disk / ~15 s cold-disk, ~710 MB resident before transport listens).
The 0.1.6 lazy-no-evict design is sound in isolation but conflicts with the current mcp-ts-core OTEL setup order. A framework-level fix (delay HTTP instrumentation, or scope it to the MCP endpoint) would unlock lazy-load; tracked separately.
57 tests pass; bun run devcheck clean.