Release 0.3.2

Added

• A UI token selector analyzes the screenshot and drops what the model does not need, which cuts screenshot tokens
• Settings persist through a settings manager with a typed store
• Escape and stop requests coordinate through dedicated coordinators, so one press ends everything that is running
• Text-to-speech shows what Juno is saying, with metadata, while it plays
• Dictation and always-on listening share one Whisper instance