RAG-5: Semantic embedding fallback for conceptual queries (stretch goal)

[nonogv]

Context

Stretch goal. Depends on #67 (RAG-2). Only worth implementing if BM25 proves insufficient in practice.

Motivation

BM25 handles keyword-heavy queries well ("Operator ratio", "Glue Compressor attack"). It struggles with conceptual queries like "how do I make it warmer?" or "why does my mix sound muddy?" — queries where the answer is in the docs but no keywords overlap.

What

If retrieveLiveDocs returns no results above the BM25 threshold, optionally embed the query using the user's already-configured provider API key:

Provider	Embedding model
OpenAI	text-embedding-3-small (~$0.02/1M tokens)
Gemini	text-embedding-004
Anthropic	No embedding API — skip, fall back to web_search

Pre-computed chunk embeddings stored in data/live-docs-embeddings.bin, generated by npm run scrape-docs --embed. At query time: embed the query, cosine-similarity against chunk embeddings in-process, return top-k above threshold.

Constraints

• Only fires when BM25 returns nothing — never adds latency to keyword-matched queries
• Uses the user's existing API key — no new credential required
• --embed flag on the scraper is opt-in; the extension works fine without embeddings
• Embeddings file is gitignored (too large) — generated locally or in CI with --embed

Decision gate

Implement only after RAG-3 ships and real usage shows BM25 gaps. Track missed queries in a separate issue before committing to this.