anomaly_mode for state classifications

***Edited/narrowed after building prototype.

Problem: Right now, State Classifications always force a "best match". This is actually the perfect behavior for 100% of current common use cases but prevents new use cases where we want the system to say, "This is something different." While anomaly can be trained into state classification. It takes data to do that. Data isn't always available.

The Solution: I am requesting zero shot, smart anomaly detection.

Proposed Configuration (for this opt-in mode only):

• threshold: A confidence cutoff. (e.g., "If the closest match is less than 90% confident, label it as anomaly").
• duration with sampling: A way to require the anomaly to last for seconds, minutes, hours before alerting a change. Also # of confirmed samples.

Use Cases Immediately Unlocked:

• IR illuminated rat trap in attic: user does not know what the snap trap is going to look like after it goes off, only that the specific area will look different.
• Anywhere that may be difficult or time consuming to access and/or no variance data is available.