ActivityPubDeliveryService::pool() and queueDelivery() only perform HTTP
delivery when app()->environment('production'); outside production they log
and return. The pool/job delivery tests were written before that guard was
added (commit 868e09b64) and never forced production, so no requests were
made and every assertSent/assertSentCount failed (6 failing tests).
Wrap the delivery-exercising tests in a forged production environment
(restored afterwards) and seed the public-ips DNS cache plus an empty
banned-domains cache so validateDestination() passes without a real DNS
lookup or DB query. Seed after factory creation, since the lazy DB refresh
can flush the cache store.
Also fix a by-value capture bug in the onError assertions: the $errors
array was captured through an arrow-fn wrapper, so the inner callback's
by-reference mutations never reached the assertion. Use a regular closure.
No production code changed; the production delivery gate is intended
behavior.
queueDelivery() runs synchronously from the v1 follow/unfollow endpoints
(via Helpers::sendSignedObject), which commit local state before delivery
and have no try/catch. After the Http::send() rewrite, a ConnectionException
from a momentarily-unreachable remote was rethrown out of queueDelivery(),
turning a best-effort delivery into a 500 for the user after the follow/
unfollow was already persisted. For unfollows, a retry then hit the
isFollowing==false branch and never re-sent the Undo, diverging state.
Treat transport failures (ConnectionException) as best-effort on this
single-delivery path: log, record host health, and return without
propagating. Other exception types (invalid sender/destination, signing,
serialization) still throw, matching pre-rewrite precondition behavior.
Also widen SendUpdateActor's per-user catch from HttpException to Throwable
so a single bad host no longer aborts a fleet-wide actor update (the old
HttpException catch is dead for ConnectionException/invalid-destination).