feat: local semantic search over cached abstracts #21

Summary

Add mosaic search --local (or mosaic semantic QUERY) to perform embedding-based semantic search over the local SQLite cache without hitting the network.

Motivation

After accumulating hundreds of cached papers, users need a way to surface semantically similar results from their personal library — not just exact keyword matches. This is especially useful for re-discovering older cached papers that share a concept but not the same keywords.

Proposed behaviour

# Semantic search over local cache
mosaic search "self-supervised contrastive learning" --local

# Explicit semantic command
mosaic semantic "methods for learning representations without labels"

# Limit to papers with PDFs downloaded
mosaic semantic "attention mechanism" --downloaded-only

Implementation notes

• Embedding model: Use a lightweight local model — sentence-transformers with all-MiniLM-L6-v2 (~22MB) is a good default, or fall back to TF-IDF cosine similarity if sentence-transformers is not installed
• Index: Build an in-memory FAISS/numpy cosine index at query time from the abstracts column; for large caches (>10k papers), persist the index to a file alongside the SQLite DB
• Extra dependency: sentence-transformers (optional; add as mosaic-search[semantic] extra)
• Storage: Store embeddings as BLOB in a new embeddings table, keyed by paper_uid; regenerate lazily when an abstract is new or updated
• Ranking: Return papers sorted by cosine similarity descending; display similarity score alongside the Rich table

Priority

Tier 3 — high value, higher complexity and dependency footprint