v3.39.3

Fixed

• Ollama chat, ad-hoc image analysis, and conversation summarization no longer send num_predict: -2 (Ollama's "fill context" sentinel). A model served by Ollama Cloud rejects it with max_tokens must be positive, got: -2, which failed every request on such a setup. The three agent-path output limits (CHAT_MODEL_MAX_TOKENS, VLM_NUM_PREDICT, SUMMARIZATION_MODEL_PREDICT) are now unset, so the option is omitted from the request and the server applies its own default. On a local Ollama server that is the same behaviour as before: -2, -1, and no value all stop at the same runner limit (verified against Ollama 0.32.13). On Ollama Cloud the model's own default applies. The video pipeline's fixed caps (VIDEO_VLM_NUM_PREDICT, VIDEO_SUMMARY_NUM_PREDICT) are unchanged. Thanks to @hruba202 for the report and the fix. (#614, #620)