Submit a task correction (active learning)

Captures a user correction to a normalised task and updates the tenant’s learned rule base.

Flow:

1. Loads the current task state
2. Applies the submitted field corrections
3. Learns rules from the before/after diff
4. Persists the updated execution state
5. Returns learning stats and rules generated

Designed for feedback loops where humans correct classification (owner, domain, stage, priority).