feat: enable verified new models for tool routing

[KillerQueen-Z]

Summary

• refresh the public catalog capability snapshot and record real structured tool-call evidence for newly listed models
• admit 11 verified new models to tool-required requests as availability fallbacks, without promoting them over trajectory-benchmarked primaries
• keep DeepSeek V4 Flash Vision Experimental and Qwen 3.8 Flash fail-closed after they did not return a valid tool call
• require every new fallback to exist in the host's live pricing/catalog map and continue respecting capability overrides and unavailable-model guards
• bump the package to 0.2.0 and rebuild the committed distribution

Validation

• npm run check — typecheck, 120 tests across 7 files, and production build passed
• decision snapshot reviewed: selected models, task classifications, and cost estimates are unchanged; only tool-required fallback chains expand
• npm pack --dry-run --json passed for @blockrun/router-core@0.2.0

Live tool-call probe

Verified structured function calls from the new Claude 5.1/5.5, Gemini 3.8 Flash, GPT-5.1/GPT-6, Grok 4.6/4.7, and MiMo 2.5 entries. Probe records contain only capability booleans and timestamps.