way-embed: add batch-embed command for multi-text single-model-load

[aaronsb]

Context

embed-suggest.sh evaluates vocabulary candidates by re-embedding description+vocabulary+candidate for each candidate term, then scoring test prompts against the augmented embedding. This requires loading the MiniLM model once per candidate.

With 61 candidates, total runtime is ~4s (dominated by repeated model loads at ~65ms each). The actual embedding computation per text is <1ms.

Proposed: way-embed batch-embed

A new command that loads the model once and embeds multiple texts in a single session:

way-embed batch-embed --model MODEL < texts.jsonl > embeddings.jsonl

Input (one JSON object per line):

{"id": "candidate_0", "text": "description vocabulary term0"}
{"id": "candidate_1", "text": "description vocabulary term1"}

Output (same format, with embedding vectors):

{"id": "candidate_0", "embedding": [0.123, ...]}
{"id": "candidate_1", "embedding": [0.123, ...]}

This would reduce embed-suggest.sh from N model loads to 1, bringing 61-candidate runtime from ~4s to ~200ms.

Why this matters

The suggest tool is the vocabulary tuning feedback loop for ways. Faster iteration means tighter feedback between "add term" → "measure impact" → "accept/reject". At 4s it's usable; at 200ms it's interactive.

The same batch-embed command would also benefit:

• Corpus generation (currently sequential in way-embed generate)
• Any future tool that needs to embed multiple texts for comparison

Implementation notes

• Model load happens once in engine_init()
• The embedding function engine_embed() is already stateless per-call
• Main work: add a loop that reads stdin JSONL, calls engine_embed() per line, writes to stdout
• No changes to existing generate or match commands