Two chunking strategies, recorded per chunk

Implement two chunking strategies - fixed-size with overlap, and semantic or section-based. Store which strategy produced each chunk, so the two can be compared later. Emit chunks/chunks.jsonl carrying chunk_strategy and token_count per chunk.

Chunking is the single highest-leverage retrieval decision, and it is invisible unless you label it. Recording the strategy now is what makes Phase 2's error analysis possible at all - otherwise you cannot tell whether a bad answer was a bad chunk.