Fix #432 regression — self-aggregating projection dispatcher dropped (2.9.5)

[jeremydmiller]

Problem

Published 2.9.3 and 2.9.4 break self-aggregating single-stream projections: at runtime they throw

No source-generated dispatcher found for SingleStreamProjection<T, TId>

This regresses a core Marten pattern (e.g. SingleStreamProjection<SimpleAggregate, Guid>) — surfaced when bumping Marten to 2.9.4 (JasperFx/marten#4714); Marten's EventSourcingTests pass on 2.9.2 and fail on 2.9.4.

Root cause

#432 (2.9.3) added a Pipeline-1 dedupe in AggregateEvolverGenerator.ExecuteCombined to avoid emitting the same evolver hintName twice when a type's Apply/Create methods are split across multiple partial declarations (CS8785). The guard reused the cross-pipeline seen set:

if (!seen.Add(SymbolKey(info.ClassSymbol))) continue;
MarkSeen(seen, info);

But MarkSeen seeds seen with the class symbol and (for partial projections) the aggregate type. So a self-aggregating type whose key was already added to seen by an earlier candidate's MarkSeen (e.g. a projection subclass aggregating that same type) had its Pipeline-1 emission falsely skipped — no evolver generated.

Fix

Dedupe Pipeline-1 emissions in a set distinct from seen, keyed by (type, identity) to match the emitted hintName granularity, while still calling MarkSeen for cross-pipeline coordination. Preserves #432's partial-split dedupe; restores emission for self-aggregating types.

Verification

• JasperFx.Events.SourceGenerator.Tests 21/21 (incl. Projections split across multiple partial record/class definitions results in duplicate hint name in source generator #432's partial-split regression test).
• Marten EventSourcingTests Aggregation 228/230 (2 pre-existing skips) and TenantPartitionedEventsTests 186/186 against a local 2.9.5 build — both RED on 2.9.4.

Bumps to 2.9.5. (Recommend yanking/superseding 2.9.3 & 2.9.4 given the regression.)

🤖 Generated with Claude Code