AI tooling fixes callback safety, inference overhead, agent gaps

Four widely-used AI libraries released updates addressing operational concerns. Pydantic AI 2.32.1 now rejects nested synchronous callbacks that could cause unsafe behavior inside agent runs. Ollama 0.32.15 (prerelease) adds model-metadata caching to cut per-request overhead. n8n 2.35.5 stops restarting slow task runners and optimizes expression-engine initialization. Transformers 5.15.1 fixes issues in DFlash candidate generation and Gemma 4 video handling.