Pin HuggingFace model revisions for reproducible benchmarks

[urrme]

The Embedder pins a HuggingFace model revision, but the reranker, SPLADE retriever, and grounding verifier call from_pretrained without a pinned revision=. If HF republishes those weights, benchmark numbers can drift silently.

Scope (good first issue):

1. Find the unpinned loads (start at src/nexusrag/retrieval/reranker.py; also SPLADE + grounding).
2. Record each model repo's current commit SHA and add revision="<sha>" to the call.
3. Note the pins in PROVENANCE.md.
4. Re-run make eval and confirm the committed benchmark numbers still reproduce.

Self-contained; no architecture changes needed.