Subgroups "2.0" and "2.1" on the connected devices had finished everything. A
ChillDKG COMPLETE for each, a birth certificate signed by the parent's quorum for
each, a key state signed by their own quorum for each -- all four events sitting in
the coordinator's database. The coordinator was offered no way to create the room.
`observeSignedGroupKeyState` resolved which key to ask about by looking the
ceremony up **from the room**:
getLatestSessionFor(chatRoomId, parentChatRoomId) // parent was null here
The view model calls it with no parent, so that reads "the ceremony in this room
that is *not* for a subgroup". A subgroup's ceremony room holds only the
subgroup's ceremony, so it matched nothing, derived no key, and reported "not
signed" over a state in the same database. The button is gated on that boolean, so
it never appeared.
This is the same mistake as `0b65d702` one layer down. That commit fixed the
*session* lookup by having the transcript name the ceremony; it left the key-state
and certificate lookups still re-deriving a ceremony from the room. A room does not
have one, and every lookup that assumes it does is wrong in a different way: by
room it finds none here, and by "newest in room" it would have found the wrong one
in the rooms from the previous report.
So both observers now take the ceremony they are about.
`observeSignedGroupKeyState(dkgSessionId)` derives the subject room from that
ceremony's own threshold key, and `observeBirthCertificate(dkgSessionId, parent)`
does the same. Neither can disagree with the ceremony the screen is showing,
because it is handed the same one.
In the view model they move into `observeForCeremony`, re-pointed when the
ceremony changes the way the message watcher already is -- previously the key-state
watcher was started once at init against a room, which is what let it drift from
the session on screen. `observeKeyState` keeps only the room-scoped watch of
signing sessions, which is genuinely a room question.
One test in `SignedGroupKeyStateTest`: a key state the group really signed is
reachable from the ceremony that produced it, and a ceremony with no key answers
null rather than throwing -- this flow runs from before a ceremony finishes.
397 common tests, 717 jvm tests, `m3Audit` meets every budget.
The two stuck subgroups will offer "create the subgroup" on the next build: their
key states are already signed and on file, and nothing about them has to be redone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
31e6fc6425