feat(models): pull OpenRouter catalog from its live models API

[Pizzaface]

Problem

OpenRouter models came from a static list: @earendil-works/pi-ai ships a build-time snapshot (providers/data/openrouter.json, 346 models), and pi overlays it with pi.dev's catalog (withRemoteCatalog). Both lag OpenRouter itself — new models never show up until an upstream release.

Change

Register a native openrouter provider that spreads pi-ai's own provider (auth — env key + OAuth — and streaming untouched) and replaces only getModels() with OpenRouter's live https://openrouter.ai/api/v1/models.

• Cached in ~/.pizzapi/openrouter-models-cache.json, 24h TTL, 0600
• Sessions warm it stale-while-revalidate (extensions/openrouter-provider.ts): cached list is used immediately, a refreshed fetch re-registers the provider so the runtime snapshot updates mid-session
• Fetch only runs when OpenRouter credentials exist (env or stored)
• Daemon listConfiguredModels() reads the cache too, so the web model picker shows live models
• Filtered to models advertising tools (same rule pi-ai's generated catalog uses); pricing converted per-token → per-million
• Any failure (offline, API down, empty/corrupt cache) falls back to the static catalog

Verification

static: 346 live: 352 fetched: 352
live-only sample: [ "meta/muse-spark-1.2-contributor", "deepseek/deepseek-v4-flash-vision-exp", "stealth/ox-alpha", "~z-ai/glm-latest", "z-ai/glm-5.3" ]
resolves live-only model: meta/muse-spark-1.2-contributor true

• packages/cli/src/openrouter-models.test.ts — 10 new tests (mapping, cost conversion, tools filter, TTL, stale refetch, failure paths, corrupt cache, both registration shapes)
• bun run typecheck clean
• bun test packages/cli/src: 2823 pass / 254 fail — identical 254 failures on origin/main baseline (pre-existing full-suite cross-file pollution), so no regressions

Skipped

• Daemon periodic refresh loop (like runner-ollama-models-cache.ts) — sessions warm the cache; add if idle runners serve >24h-stale lists
• Registration in provider-auth.ts / commit-message-generator.ts — auth behaviour is identical since the provider spreads pi-ai's; add if either needs live-only model ids
• Docs — no new flag, env var, or config key