[jeremydmiller]

The 6.x prep for Wolverine 7.0 removing MultipleHandlerBehavior. Each bullet is its own commit.

1. ClassicCombineIntoOneLogicalHandler obsolete, plus a startup warning and AssertNoCombinedHandlers()

• [Obsolete] on the enum value, with a message that points at Separated and the 7.0 change. It is still the default. The repo's global NoWarn already includes 0618, so nothing in the repo needed pragmas.
• When an app actually has message types whose handlers come from more than one class, Wolverine logs one warning at startup that lists them. Several methods on one class, like a saga's Start/Handle, don't count.
• New opts.AssertNoCombinedHandlers() (on WolverineOptions, next to AssertNoBatchHandlerConflicts()) turns that warning into a startup failure for CI. Added a short docs section in handlers/index.md.

2. Latent Separated bugs: these consumers read HandlerGraph.Chains, so they never saw per-endpoint chains (new internal HandlerGraph.EveryChain() / SagaChains())

• Requeue-policy listener validation (Pulsar hot-tail listeners silently drop deferred messages in every mode #4060), and the failure-rules section of describe.
• MessagePartitioning.UseInferredMessageGrouping() / MaybeInferGrouping.
• HandlerGraph.ConfigureHandlerForMessage() now also applies to every ByEndpoint chain, sticky ones included.
• describe-routing <message> now lists the separated handlers along with their endpoints.
• The saga store diagnostics (EF Core, RavenDB, RDBMS, Marten) and the CosmosDB saga validator.
• Bobcat HandlerWarmUp compiles the per-endpoint chains through HandlerFor(type, endpoint).

3. Per-handler OTel tag. TagHandlerPolicy now captures the distinct handler type names when the policy runs. Before, the frame re-checked the call count during codegen, after SagaChain had already cleared its calls, so no saga ever got handler.type. A combined chain is now tagged with all of its handler types. Regenerated the committed Wolverine.AotSmoke.Saga output (it gains the two tags). I ran codegen write on all 11 drift-gate projects, and only those two files changed.

4. Per-chain dedup scoping. When a message type has more than one running chain, each chain claims id + "|h:<fnv1a of handler type names>". Before, the first claim won and the other handlers dropped the message. A message type with a single chain keeps storing its ids unchanged. I used a short hash instead of the type name because the id column is varchar(250).

Tests run locally (net9.0, Release)

• Full wolverine.slnx build with -f net9.0: 0 errors. The only warnings are the existing NU1902 MessagePack ones. The first two passes hit a transient sourcelink race in WolverineWebApiFSharp, which builds cleanly by itself, and the third pass was clean.
• CoreTests full suite: 3832 passed, 8 skipped, 0 failed. The new tests are combined_handlers_warning (5), separated_handler_chains_are_not_skipped + separated_saga_chains_are_not_skipped (8) and handler_type_activity_tagging* (6).
• MessageRoutingTests: 25/25.
• PostgresqlTests: the new deduplication_with_separated_handlers and logical_message_deduplication, 16/16.
• Wolverine.Bobcat.Tests: 269/270 on the first run. The one failure was a Polecat TrackedSession timeout in polecat_consistency_boundaries, which passed 2/2 on rerun. The new warming_separated_handlers passed.
• The new tests for inferred grouping, OTel tagging, Bobcat warm-up and dedup were confirmed red before their fixes.

Closes #5133

🤖 Generated with Claude Code

https://claude.ai/code/session_01DCggXgxnG3EBXoR9rrq37U