docs: add approval-bound action metadata example

[jw-ond]

Summary

Adds a worked example to the Interrupts documentation showing how sensitive tool-call approvals can be bound to a proposed action fingerprint.

The new section uses existing AG-UI concepts only:

• interrupt metadata carries the proposed action fingerprint and policy metadata
• resume payload echoes the approved fingerprint and can carry a verifier receipt
• the enforcing agent or server recomputes the actual execution fingerprint before emitting ToolCallResult
• action drift should terminate with RunError instead of a successful result

This does not add new protocol fields or governance semantics to AG-UI. It documents a practical pattern using metadata and resume payloads that are already part of the interrupt lifecycle.

Notes

Running Prettier on docs/concepts/interrupts.mdx also rewrapped existing prose and tables in the same file. The substantive addition is the Approval-bound action metadata section.

Test plan

• npx prettier --check docs/concepts/interrupts.mdx
• git diff --check