Adom Bridge

Reproduced from winvm with your exact shape (scenario reminder, buttons as objects, no wait): ab answers displayed, and NOTHING is mirrored - no MIRROR line on the laptop, no card. Cause, not a 2.1.93/94 regression: notifications.rs mirrors a buttoned toast as a notify-decision card only when wait:true (decision_prompt = payload.wait && !buttons.is_empty()), and mirrors a plain toast only when it has NO buttons - so a buttoned toast raised without wait mirrors nowhere. Every proof that 'worked' (the 2026-09-09 circuit, my 2.1.94 proof) passed wait:true. Workaround for your harness today: add wait:true (and a timeoutSeconds). Fix coming as 2.1.96: a buttoned toast is registered and mirrored as a decision card whether or not the caller waits; a decision made in Hydrogen or on another box records the notify_response exactly as a click on the origin box does, so a polling caller sees it.

Fixed in 2.1.96 (insiders; winvm canary, then ConfRoomROG and AdomLapper). Cause was the GATE, not the transport: decision_prompt was payload.wait && !buttons.is_empty(), and the plain-toast mirror skips any toast that HAS buttons, so wait:false + buttons entered no mirror at all while ab still answered action:displayed. It is now !buttons.is_empty(), and that Vec is built from BOTH request shapes (structured buttons:[{label}] and simple actions:[label]), so your three calls and my repro take the same path. Second flaw caught before shipping: the teardown (forget_notify_decision + permission_mirror::dismiss + shell-approval-dismiss) sits immediately after the wait loop, so simply widening the gate would have mirrored the card and dismissed it in the same instant. The teardown is now the waiting path's; a non-waiting prompt gets a watcher after its result is on the wire that polls has_notify_response (non-destructive, so the answer stays for your notify_response poll) up to the card lifetime, then retires every copy. A non-waiting card advertises 600 s instead of 60 s, matching the toast that sits in the notification centre. Proof on ConfRoomROG (ab 2.1.96, ah 1.0.322), image attached: your shape 1 (local, scenario reminder, buttons objects, no wait) and shape 2 (same from winvm) BOTH render as decision cards in Hydrogen, with 1.0.322's badges - orange WI cloud for winvm, purple CR laptop for the local one - and Hydrogen's header reads 2 approvals waiting. Note for your harness: hd_eval is dev-only in the public build (verb_moved_to_hd_dev); hd_api is the pass-through, and ah's control API has no /permission/state, so DOM counting needs the hd-dev build or a screenshot.

Second half proven: the non-waiting cards are not immortal. Same two cards, same Hydrogen, 11 minutes later (raised 10:30:39 and 10:30:52 with the 600 s life, screenshot at 10:41:40): both gone, Hydrogen's '2 approvals waiting' chip cleared, and ab's own permission stack on ConfRoomROG reports no repro182 card. So the watcher retires every copy - ab's stack, Hydrogen, the peers - when the lifetime runs out, exactly as the waiting path does when it times out. Fleet is on 2.1.96 (winvm canary through the real embedded update path, then ConfRoomROG and AdomLapper); Linux CLI 2.1.96 published and gallia's relay module refreshed. Closing once the smoke suite reports clean.

Closing: 2.1.96 insiders on the whole fleet, both halves proven on ConfRoomROG with ah 1.0.322 (cards appear for non-waiting buttoned toasts, local and cross-box, with the right badges; and they retire everywhere when the lifetime runs out); smoke suite on 2.1.96: 60/60 executed, 0 unexpected failures.