Embedding model cold-loads on every prompt — add monolingual/bilingual language mode

[aaronsb]

Problem

Every UserPromptSubmit hook invocation spawns way-embed match as a subprocess — up to twice per prompt:

1. EN corpus (21 MB GGUF, 132 entries)
2. Multilingual corpus (126 MB GGUF, 1,411 entries)

Each invocation cold-loads the full GGUF model into memory, computes the query embedding, scores all corpus entries, then exits. The 126 MB multilingual model load causes a visible CPU spike (~0.4s, fans spin up) on every single user message.

The corpus itself is not rebuilding — corpus --if-stale correctly exits in ~1ms when the manifest is fresh. The cost is purely model loading at inference time.

Evidence

# Per-prompt cost (ways scan prompt):
0.35s user, 0.04s system — 103% cpu, 0.38s wall

# Model sizes:
minilm-l6-v2.gguf                      21 MB   (EN)
multilingual-minilm-l12-v2-q8.gguf    126 MB   (multilingual)

# Corpus sizes:
ways-corpus-en.jsonl                   132 entries
ways-corpus-multi.jsonl              1,411 entries

Call path

UserPromptSubmit hook
  → check-prompt.sh
    → ways scan prompt --query "..."
      → batch_embed_score()          [scoring.rs]
        → run_embed_match(en_corpus, en_model)       # spawn way-embed, load 21MB
        → run_embed_match(multi_corpus, multi_model)  # spawn way-embed, load 126MB

Proposal: Language mode config

Add a configuration property that controls which embedding models are loaded:

bilingual        → EN + multilingual models (current default, ~147 MB loaded per prompt)
monolingual:en   → EN model only (~21 MB per prompt)
monolingual:de   → multilingual model only, filtered to de corpus entries

Benefits

• Monolingual EN users (majority case) skip the 126 MB multilingual model entirely — ~6x less memory, faster inference
• Monolingual non-EN users skip the EN model, use only the multilingual model with their language's corpus subset
• Bilingual preserves current behavior for users who actually need it
• No cross-platform complexity (a daemon/socket approach would diverge across OS)

Where it lives

The setting could go in:

• settings.json under a ways key (e.g., "ways": {"language_mode": "monolingual:en"})
• Or as a frontmatter default in hooks/ways/core.md

Implementation sketch

1. Add language_mode config reading to batch_embed_score() in scoring.rs
2. Skip corpus/model pairs that don't match the configured mode
3. Default to monolingual:en (matches the majority use case; bilingual users opt in)
4. ways status should report the active language mode

Future consideration

A longer-term fix would keep way-embed resident (daemon with socket/pipe interface) so the model loads once per session rather than per prompt. However, this introduces cross-platform complexity (Unix sockets vs named pipes vs TCP) and process lifecycle management. The language mode approach is simpler and addresses the immediate pain.