Default bundled router model requires sentence-transformers, but it isn't a declared dependency

[rogercloud]

Repro

pip install xrouter-llm==0.1.1
# build the default RoutingService from default_model_path(), then call route()

The default predictor returned by default_model_path() uses a SentenceTransformer embedding backend. On the first route() call the encoder lazily does from sentence_transformers import SentenceTransformer (encoders.py), which raises:

ModuleNotFoundError: No module named 'sentence_transformers'

Cause

sentence-transformers (and its torch dependency) is not declared anywhere in xrouter-llm's dependencies — the package only requires huggingface-hub, joblib, numpy, pandas, pyyaml, scikit-learn, scipy, and exposes only a dev extra. So the bundled default model cannot actually run after a clean install.

Impact

Any downstream consumer following the documented install gets a router that throws on the first real route.

Suggested fix (any of)

1. Declare sentence-transformers as a runtime dependency; or
2. Provide an extra (e.g. xrouter-llm[st]) and document that the default model needs it; or
3. Ship a default model that uses the TfidfSvdEncoder (no torch).