fix(retrieval): build index with the caller's config, not the default config.yaml

[CGFixIT]

Summary

retrieval/indexer.py::build_index(config_path) loads its configuration from config_path and forwards it everywhere except the embedding call:

clean_chunk = sanitize_chunk(chunk, config_path)      # honours config_path
...
batch_embeddings = get_embeddings_batch(batch_chunks)  # ← uses default config.yaml

get_embeddings_batch(texts, config_path="config.yaml") accepts a config_path but build_index never passed it, so the semantic index was always embedded with the model named in the default config.yaml, regardless of the config the caller actually supplied.

Why it matters

Query-time embeddings already honour config_path:

# retrieval/hybrid_search.py
def semantic_search(self, query, ...):
    emb = get_embedding(query, self.config_path)

So with a non-default config that selects a different embedding model (or dimension), the corpus vectors and the query vectors come from different models — they're no longer comparable, and semantic retrieval silently degrades or breaks (mismatched dimensions, or meaningless cosine scores). It also undermines test isolation, where a tmp_path config is the whole point of pointing the indexer at a throwaway model/index.

The fix

One line — forward the already-available config_path:

batch_embeddings = get_embeddings_batch(batch_chunks, config_path)

Index-time and query-time embeddings now always come from the same model.

Tests

Adds TestBuildIndexConfigPropagation::test_config_path_reaches_embeddings, which mocks get_embeddings_batch + chromadb and asserts build_index forwards its config_path. Verified it fails on the old code and passes with the fix.

tests/test_indexer.py ........  8 passed

🤖 Generated with Claude Code

https://claude.ai/code/session_01LvLWMML8cpBq2q81kL1ByJ

---

Generated by Claude Code