Smarter reasoning support and refusal fallbacks

A round of model-layer upgrades focused on reasoning correctness and graceful failure.

• Catalog-driven reasoning. Whether a model supports extended thinking, and which controls it accepts (effort levels, token budgets, or both), is now read from live model-catalog metadata instead of guessed from model names. Around twenty models newly expose reasoning correctly, including the Grok 4.x, DeepSeek V4, and Qwen 3 families.
• Adaptive thinking on Claude 4.6+. Claude Sonnet 5, Fable 5, Mythos 5, and Opus 4.7+ require Anthropic's adaptive thinking API; Runtype now selects it automatically. This fixes reasoning-enabled agents on Claude 5 models, which previously could fail without output. Older Claude models keep budget-based thinking.
• Refusals hit your fallback chain. When a model's safety classifier refuses a request, the step no longer returns a silent empty result: any fallback chain with the default error trigger takes over, skipping pointless retries of the same model and moving to a different model or your configured fallback message. If the whole chain refuses, the step fails loudly instead of succeeding empty.
• Mixlayer vision. Mixlayer's Qwen 3.5 and 3.6 models now accept image input.
• Gemini 3.6 Flash for new accounts. The fast-model slot for new accounts moved from Gemini 3.5 Flash to 3.6 Flash: same 1M context and capabilities, cheaper output. Existing accounts are unchanged.