GH-4595: stop the protocol version header surviving a round trip

[jeremydmiller]

Closes #4595.

RedisEnvelopeMapper wrote wolverine-wolverine-protocol-version twice whenever it re-sent an envelope it had received. The round trip is the cause, not the Redis mapper: writeIncomingHeaders copies every wire header into Envelope.Headers, so the received envelope kept wolverine-protocol-version as a header of its own, and MapEnvelopeToOutgoing stamps its own copy onto every send. Second pass, two fields with the same name.

Where it bites

Redis's outgoing headers are a List<NameValueEntry> and Kafka's are a Headers list — both accept the same name twice. Everywhere else the outgoing headers are a dictionary, so the second write collapsed onto the first and nothing showed. So the visible damage is a malformed stream entry, most easily reached through the native dead letter queue under BufferedInMemory or Durable listening (those route the failure through TryBuildDeadLetterSender → the mapper, rather than through RedisStreamListener.MoveToErrorsAsync, which writes its fields explicitly). A consumer reading that entry into a keyed structure throws:

System.ArgumentException : An item with the same key has already been added.
Key: wolverine-wolverine-protocol-version

This is latent on main only because the Redis DeadLetterQueueTests currently listen Inline. It surfaced in CI on #4506, which is not its cause.

The fix

One line, in EnvelopeMapper.MapIncomingToEnvelope: drop the protocol version from Envelope.Headers as the envelope comes in.

Central rather than per-transport, because the exposure is not Redis's. Nothing in the codebase reads the protocol version off Envelope.Headers — it is plumbing that the outgoing mapper re-stamps on every send — and every mapper that copies wire headers wholesale keeps it. Kafka had the identical duplicate; the test proves both.

Tests

envelope_mapper_round_trip_4595 in each of the Redis and Kafka test suites. No broker: they map an envelope out, read it back, and map it out again, exactly as the dead letter sender does. Red before the change (3 of 4 Redis, 2 of 2 Kafka), green after, and red again with the one line commented out.

Clean dotnet build wolverine.slnx -c Release -f net9.0.

Not fixed here

The issue's closing observation stands and is worth its own issue: the two Redis dead-letter paths write different field shapes — MoveToErrorsAsync writes exception-type / exception-message / failed-at and friends, the dead letter sender path writes mapper-shaped wolverine-* fields. Whether native dead letter metadata should depend on the listener mode at all is the larger question behind this bug, and is not something to settle inside a one-line fix.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VDUrBeB4tTnKj4AExCS1nj