feat: enrich model metadata via models.dev with local cache

[ricatix]

Summary

Enriches model entries with real metadata from models.dev instead of injecting empty {} templates.

What changed

• src/cache.ts — 24-hour TTL cache at ~/.cache/opencode-9router-plugin/models.dev.json, atomic write via temp+rename
• src/models-dev.ts — Fetches models.dev/api.json, builds flat lookup index by model name across all providers. Cache-first, 15s fetch timeout, graceful fallback
• **src/model-mapper.ts\** — Maps models.dev fields to OpenCode schema (name, family, cost, context limits, modalities, capabilities). Minimal template fallback for combo models (no /`) and unmatched models
• src/index.ts — Uses resolveModel() instead of ||= {} for each discovered model
• src/cli.ts — Shows models.dev cache age in check command output

Behavior

Model ID type	Example	Result
Provider model (/ present)	mistral/mistral-large-latest	Full metadata: 262k ctx, cost, vision+tool_call
Combo model (no /)	my-custom-combo	Minimal template
Unknown model (/ present, no match)	unknown/some-model	Minimal template
models.dev unreachable	any	Minimal template (graceful fallback)

Tested

• openai/gpt-4o → 128k ctx, cost, text+image+pdf ✓
• anthropic/claude-sonnet-4-20250514 → 200k ctx, cost, text+image+pdf ✓
• google/gemini-2.5-flash → 1M ctx, cost, text+image+audio+video+pdf ✓
• deepseek/deepseek-chat → 1M ctx, cost, text ✓
• groq/llama-4-scout-17b-16e-instruct → 128k ctx, cost, text+image ✓
• Combo models → template fallback ✓
• Cache: 2.2MB, created on first access, TTL 24h ✓