M3-HYDE — Vector retrieval (HyDE) mode

Three-stage retrieval flow. (1) A small independent generator model (default qwen2.5:3b) writes a 1-3 sentence hypothetical answer in the style of a real archival answer — naming the value, row, column. The generator does not know the real answer; the hypothetical is purely a search query. (2) The hypothetical answer is embedded with `nomic-embed-text` (open-weight, 768-dim, ~270 MB, runs locally via Ollama). (3) Cosine similarity against the pre-built card index (`evaluation_runs/hyde/card_index.jsonl`, 407 entries, ~6.5 MB) — top-1 card is then served to the evaluation model under the existing M3-L4 prompt. The evaluation model never sees the hypothetical answer; only the retrieved card reaches it. Key advantage over M3-IDX two-shot: a 3B local model can do the retrieval for any size evaluator, including ClimateGPT-13B whose 4K context can't hold the V27/V35 indexes.