SemanticChunker strategy (depends on #12)

[nleroy917]

Summary

Add a SemanticChunker strategy to the chunkers module introduced in #12.

Depends on #12 (the chunkers/ module + Chunker ABC + registry must exist first). This issue tracks the specific semantic strategy; it was the original motivation for splitting chunking out of the model layer.

What

Instead of slicing text into fixed N-token windows (FixedTokenChunker), a semantic chunker splits on meaning boundaries — keeping semantically coherent spans together rather than cutting mid-thought at an arbitrary token count.

Typical approaches to evaluate:

• Sentence/paragraph grouping — split into sentences, then greedily pack sentences into a chunk until the token budget is hit (cheap, no model).
• Embedding-similarity boundaries — embed sentences, start a new chunk where adjacent-sentence cosine similarity drops below a threshold (the "semantic chunking" people usually mean). Costs an extra embedding pass.

Design fit

SemanticChunker is just another strategy behind the same Chunker ABC (chunk(text) -> list[str]) and registry, parameterized by the same max_tokens budget (= min across configured models) so its output still fits every model. Selected via the chunking: config block:

chunking:
  strategy: semantic
  max_tokens: 256        # optional override of the derived budget
  # similarity_threshold: 0.6   # for the embedding-similarity variant

Open questions

• Sentence-grouping (no model) vs embedding-similarity (extra pass) — start with the cheaper one?
• For the embedding-similarity variant: which model embeds the sentences (a configured one vs a small dedicated one), and does that reintroduce model coupling we just removed?
• Sentence splitter dependency (spaCy / nltk / a lightweight regex splitter).