Thinking fixes: strip client max_tokens (enforce n-predict 2048), deepseek reasoning-format

llama-swap dual-specialist setup: config.yaml