Interruption recovery, part 2: re-auth and resume, late-load re-resolve, focus-steal check, learn prompt, real-app and chaos proof #1903

Part 1 is #1902: provably read-only dismissal of known interruptions. Remaining, in this order:

1. The focus-steal race (safety). Add a post-action continuity check: the window or element that received the input must be the one verified (focus/foreground right after dispatch, plus a frame diff of the target region). On a mismatch, halt and reconcile; never retry. Test it with a timed injected popup.
2. Pixel and desktop classification (safety). Classify the whole dialog text, title and owning window against an allowlist of system interruption classes. Anything that mentions record content, a patient, or save/apply/change is never benign. If the class can't be proven, halt.
3. Session expiry. Re-authenticate (secrets via Sign-in mid-batch: session expiry, MFA and SSO through the OS keychain or a vault, never in logs #1893 when it lands), resume at the last verified checkpoint, and reconcile, never replay, the write that was in flight.
4. Late page loads. A bounded re-resolve with an identity re-check; never a blind click retry.
5. The learn prompt in the halt UI.
6. The real-app proof on the zen lab's OpenEMR and Dolibarr, then the same proof behind the shared chaos proxy (--chaos vpn-bad).
   Pre-register the cases.

Code pointers and briefs from part 1 are in the coordinator's notes (interruption-recovery.md). Ask for it in the inbox if you need it, or read #1902's description. Queue position: after the #1901/#1899 gates, per the queue freeze; items 1 and 2 may go earlier, since they're safety fixes to #1902.