[jeremydmiller]

Closes #1101.

A message published with slice.PublishMessage(message) went out with its tenant alone, even though the slice held the correlation id of the event that caused it. Under the async daemon the bus then fell back to one id per page, so two unrelated flows landing in one page shared a correlation id. Inline, each flow lost its id.

What changes

When a message has no explicit metadata, its slice now supplies the correlation id if every appended event in the slice agrees on one. This applies on the async path and both inline paths, which all go through one method on EventSlice, so they can't drift apart.

• No agreement means no change. A null on an appended event, a disagreement or an empty slice makes exactly the old tenant-only call. Taking the last event's value would be a guess, because a slice can span several flows.
• Stand-in events don't count. slice.Reference<T>() and composite stages add events nobody appended, and their null ids say nothing about the flow. Counting them would have left every enriched slice with the bug. References<T> (and so Updated<T>) and ProjectionDeleted<TDoc, TId> now carry an internal ISyntheticEvent marker.
• MessageMetadata.From(IEvent) is new, so the mixed case is one line: slice.PublishMessage(msg, MessageMetadata.From(e)). It takes the event's tenant and correlation id.
• slice.PublishMessage(message, metadata) still wins, and the order between the two lists is unchanged.

Decisions worth checking

• This is the default, with no option. An opt-in is the state the issue says is the problem.
• Causation is deliberately left out, on both the default path and From. The issue suggested IEvent.CausationId, but it does damage here. Wolverine sets ConversationId = originalEnvelope.Id while handling a message, so today a side effect handler's events correctly name that side effect as their cause. A causation-id header takes precedence over that in OutboxedSessionFactory, so copying the event's causation would make those events name the root message instead, and flatten the chain. MessageMetadata.CausationId stays available to anyone who sets it on purpose. A test pins this.

Release note

With event metadata on, side effects from slice.PublishMessage(message) now carry the correlation id that all of the slice's appended events agree on. The id used to be the daemon page's id (async) or a fresh one (inline). This changes the envelope's correlation id for existing side effects. A custom IMessageSink or IProjectionBatch that overrides the metadata overload now receives these calls instead of the tenant-only overload.

With event metadata off, or when nothing agrees, nothing changes. New public MessageMetadata.From(IEvent).

Known limits

• Still falls back: fan-out and ReplaceEvent slices, because CloneEventWithNewData doesn't copy CorrelationId, despite its doc comment. Fixing that may change what stores write on projected documents, so it needs its own issue.
• Composite stages: a stage holding only Updated<T> events has no appended event to read, so its side effects fall back too.
• Wolverine half: with event metadata off, the page-wide id still merges flows. That belongs in Wolverine.

Testing

Full gate on this branch, locally on net9.0 and net10.0: 7,788 passed, 0 failed, 62 skipped. With the call sites or the stand-in skip reverted, the new tests failed, and they also failed against a "take the last value" mutation.

That was the third gate run on this commit. The first two each failed on a test this change doesn't touch:

• CommandLineTests.DescribeCommandTests.write_file_smoke_test_as_htm: Collection was modified inside AnsiConsole.ExportHtml, 2 of 12 in a loop. CommandLineTests has no reference to JasperFx.Events.
• SideEffectGateWarmupThrottleTests.the_gate_flip_is_driven_by_committed_progression_not_by_what_was_loaded, at a load average around 60. By reading the code it's a race between LastCommitted and the gate flip, but that wasn't reproduced.

Marten, Polecat and Fisher need nothing: each IProjectionBatch overrides the metadata overload and forwards it, and all three Wolverine bridges map the correlation id. That comes from reading their source. This has not been run end to end against a store or Wolverine, and no compliance fact proves that a store forwards metadata.

🤖 Generated with Claude Code