The first mutation (instance deploy) to a site pair whose sticky node was down
returned OUTCOME_UNKNOWN: SitePairChannelProvider.ExecuteAsync treated every
Unavailable on a non-retry-safe command as possibly-applied unless the server
sent the not-dispatched trailer, flipped stickiness, and threw. The retry then
succeeded on the live peer.
A connect-phase failure provably never sent the request, so it is now treated
like the pre-dispatch rejection: GrpcRetrySafety.WasNeverSent recognises
Grpc.Net.Client's subchannel connect failure ("Error connecting to subchannel."
with a SocketException cause) and an HttpRequestException name-resolution /
connection error with a connect-only socket error. Anything that may have been
delivered (reset, premature end, GOAWAY, server Unavailable) stays
OUTCOME_UNKNOWN and is not replayed. The site->central reconcile path uses the
same ProvablyNotApplied rule.
Claude-Session: https://claude.ai/code/session_01GT67ACmpMR7b9LGRPLijMi
Joseph Doherty
fd80aeab57