Conversational fallback with memory; reject zero-amount expenses

Dashboard chat routes through the feature registry - log expenses from the web too

Add torchvision (Qwen2.5-VL needs it); stop leaking internals to users

Warn loudly when a dataset repo is set without a token

Stop logging the bot token; redact Telegram errors

Extract dashboard auth boundary; docs

Username access, web dashboard and coach chat

Let Gradio own the server so ZeroGPU can instrument it

Run a local model on ZeroGPU instead of paid inference

Switch to Gradio SDK for free ZeroGPU tier

Deploy coco-finbot