Load sentence-transformers on the existing cache miss and construct the
preloaded model off the shared background event loop. Start warm-up before
chat restoration inside extensible init_a0, preserving start-hook control
and sharing the loaded model with ordinary runtime requests.
Remove unused speech preload imports and dead warm-up functions. Seed APT
indexes only when absent, skipping the refresh on normal container boots.
Keep the existing Python tunnel service and its provider lifecycle.
Validation: 49 focused framework tests passed, including extension-hook
short-circuit behavior, concurrent embedding-cache reuse, APT cache states,
tunnel behavior and CSRF origins. Three cold Fable chat trials per version:
median submit-to-first-token 9.588s -> 4.234s;
local preparation 6.904s -> 1.197s;
HTTP readiness 42.637s -> 41.139s;
boot-to-first-token 52.093s -> 46.657s.
Provider timing varies; these are local repeated measurements. The live
baseline already omitted APT refresh, so it receives no timing credit.
Raise the runtime container soft nofile limit before supervisord starts so WebUI and managed services inherit a larger descriptor allowance.
Add an explicit compose nofile example, document the startup contract, and cover the limit raise and hard-limit cap with focused regression tests.
