Turn an observed pattern into a policy rule

Writes a rule into the named agent's action_approval_policy. Human users only, and only when the organisation is in enforce mode.

The rule covers only what was actually approved. Five approved $5 refunds to one customer produce { refund.create, asks above $10, that recipient } — a $49 request does not match it and still reaches a human.

Refused when: fewer than five consecutive approvals; any past rejection on the bucket (a long recent run must not hide a history of refusals); the bucket is for a recipient never paid before; the amount band has no upper edge; or the action grants or destroys authority.

widen_to_action_type drops the amount and recipient constraints. It is never the default and should be an explicit choice in your UI, not a checkbox someone skims past.

The written rule is marked promoted_from_learning so an operator can tell it apart from one they wrote, and it replaces any existing rule for the same action type — appending would leave two rules where only the first ever applies.