Create multilingual way stubs for core languages

[aaronsb]

Context

PR #101 shipped dual-model embedding support. The infrastructure is complete:

• embed_model frontmatter field routes ways to EN or multilingual model
• Dual corpus generation (ways-corpus-en.jsonl + ways-corpus-multi.jsonl)
• Scanner queries both corpora and merges results
• languages.json defines 52 supported languages
• test-multilingual.sh validates cross-language similarity
• ways language shows coverage report

What's missing: actual language stubs. All 85+ ways are English-only (embed_model: en). The multilingual corpus is empty.

Test data from this session

Same-language matching with native stubs dramatically outperforms cross-language:

Lang	EN model × EN desc	Multi model × EN desc	Multi model × native desc
en	0.76	0.68	0.68
ja	-0.03	0.69	0.93
ko	-0.02	0.72	0.86
zh	-0.03	0.59	0.89
de	0.08	0.62	0.82
ru	0.03	0.76	0.85
ar	0.04	0.40	0.96
el	-0.01	0.59	0.81

Full report: docs/architecture/system/multilingual-model-evaluation.md

Stub format

Each stub is ~5 lines of frontmatter, no body:

---
description: セキュリティ脆弱性スキャンと監査
vocabulary: セキュリティ 脆弱性 CVE 監査 認証 暗号化
embed_model: multilingual
embed_threshold: 0.25
---

File naming: {wayname}.{lang}.md (e.g., security.ja.md)

Recommended starting languages

Cover the architecture tiers with 4 languages:

Language	Code	BM25	Why
Japanese	ja	impossible	CJK, highest same-language delta (+0.96), large user base
German	de	available	BM25 compound noun challenges, tests both engines
Spanish	es	available	Romance language, clean BM25 fit, large user base
Arabic	ar	impossible	Non-concatenative morphology, highest same-language score (0.96)

Team execution plan

This is ideal for an agent team (4 teammates, one per language):

Lead: Creates task list, validates dual corpus after stubs land, runs test-multilingual.sh

Per-language teammate:

1. For each of 85+ ways, create {wayname}.{lang}.md with:
   • description: — natural expression in the target language (not mechanical translation)
   • vocabulary: — domain terms in the target language
   • embed_model: multilingual
   • embed_threshold: — tuned per-way (start at 0.25, adjust based on test results)
2. Run ways corpus to regenerate dual corpora
3. Run bash tools/way-embed/test-multilingual.sh to validate
4. Run ways language --filter {lang} to verify coverage

Reviewer: Verify native descriptions are natural, not mechanical translations

Acceptance criteria

• All 85+ ways have stubs for each target language
• ways-corpus-multi.jsonl is non-empty after ways corpus
• test-multilingual.sh passes for all target languages
• ways language shows correct coverage
• Per-way embed_threshold tuned (no false positives, reasonable recall)
• ways lint --global passes with no errors

References

• ADR-107 (rewritten): docs/architecture/system/ADR-107-way-match-corpus-batch-mode-and-locale-support.md
• Language guide: docs/hooks-and-ways/languages.md
• Evaluation report: docs/architecture/system/multilingual-model-evaluation.md
• Test suite: tools/way-embed/test-multilingual.sh