processMessage legitimately returns (nil, nil) when a message is still
segmented-incomplete or gets re-queued because its hash ratchet key has
not arrived yet. processQueuedHashRatchetMessages dereferenced that nil
response (the shadowed r), crashing the whole app with SIGSEGV. Skip
such messages and leave them in the queue.
Reliability.Stop() destroyed the SDS reliability manager while Start() only
rebuilt the mvds datasync node, so the first offline->online transition (driven
by Core.connectionChanged) left SDS nil for the rest of the process and every
subsequent message — live or fetched from a store node — arrived still
SDS-wrapped, failed to decode at the application layer and surfaced as type
UNKNOWN. Stop() now tears down only the datasync node and preserves SDS, whose bloom filter and causal history are exactly the state needed to detect what was missed while offline; a new Close() releases it on shutdown, and Start()
rebuilds it if it is ever missing. Two related bugs are fixed alongside:
UnwrapPayloadFromSDS now returns ErrSDSManagerUnavailable when the manager is gone instead of silently passing the wrapped payload through (it still passes through, error-free, when a payload is simply not SDS-wrapped), and the
processor propagates that error so the envelope is retried rather than
confirmed as processed. sdsManager is stored in an atomic.Pointer, since the
hot path read it without holding the lock that Start/Stop write under.
Fixes#7363
Enables the SDS wrapping flag.
Sets up the handler that wraps SDS messages with retrieval hints. Those hints are the envelope IDs of the messages that were sent and received
Sets up the unwrapping and fetching when there are missed messages detected.
Adds a new function that enables fetching per envelope ID instead than per topic.
Andrey Bocharnikov