t6 sentiment: finbert serving behind a protocol, POST /v1/sentiment off the event loop

[vlobus]

what

• domain/sentiment.py — SentimentModel Protocol (the serving seam), pure to_score mapping (argmax label, quantized Decimal probs, rejects wrong label sets / out-of-range values), FinBertModel adapter (ProsusAI/finbert, heavy imports deferred to construction, loaded once).
• serving/ — FastAPI app factory with DI via app.state, POST /v1/sentiment (label + probs + model_version), GET /health/live. Inference runs on a dedicated single-worker executor: keeps the event loop free, serializes forward passes (no CPU oversubscription), and sidesteps HF fast-tokenizer thread-safety.
• main.py — composition root: build FinBERT once, serve with uvicorn (python -m hodlin_recommend.main).
• deps: fastapi, transformers 5.13, torch 2.12, uvicorn.

tests

• default suite fully offline — a regression guard asserts torch/transformers are never imported.
• event-gated concurrency test proves the loop serves a liveness ping while an inference is provably in flight (no wall-clock flakiness).
• opt-in real-model smoke (HODLIN_TEST_FINBERT=1): verified locally — positive/negative texts score correctly through the endpoint, model_version finbert:ProsusAI/finbert.
• full gate green: ruff, mypy strict, import-linter, pytest (67 passed).