fix(indexer): re-embed unchanged files when indexing_params change

[georgeh0]

Summary

• INDEXING_EMBED_PARAMS is now declared with detect_change=True. Before, editing embedding.indexing_params (e.g. adding prompt_name: passage) and restarting the daemon left files with unchanged content on their memoized process_file results, embedded with the old params. This affected sentence-transformers. LiteLLM was already covered because its constructor kwargs are part of the embedder memo key.
• QUERY_EMBED_PARAMS stays as is: it is read only at query time, outside any memoized function.
• Tests: StubEmbedder in conftest now records the kwargs of each embed() call, replacing the duplicate _KwargRecordingEmbedder. The registry index helpers moved from the chunking test into conftest. A new end-to-end test, tests/test_reindex_on_embed_params_change.py, covers the change.

Release note: memo entries written before this change don't record the params dependency. Upgrading from v0.2.42 is still correct, because #299 added a process_file argument and so every file re-runs once. I checked this by indexing with v0.2.42 and then with this branch. This must ship in the same release as #299.

Test plan

• New test: adding, changing, or removing a param reprocesses both unchanged files, and the embedder sees only the new kwargs. Equal params across a restart, in either key order, leave every file unchanged. The change cases fail without the fix: 0 reprocessed, 2 unchanged.
• uv run pytest tests/: 329 passed. mypy src/ tests/, ruff check, and ruff format --check are clean.

🤖 Generated with Claude Code