Stratara.ServiceDefaults 3.2.3

• net10.0

  • Destructurama.Attributed (>= 5.3.0)
  • Microsoft.Extensions.Http.Resilience (>= 10.6.0)
  • Microsoft.Extensions.ServiceDiscovery (>= 10.6.0)
  • OpenTelemetry.Exporter.OpenTelemetryProtocol (>= 1.15.3)
  • OpenTelemetry.Extensions.Hosting (>= 1.15.3)
  • OpenTelemetry.Instrumentation.EntityFrameworkCore (>= 1.15.1-beta.1)
  • OpenTelemetry.Instrumentation.Http (>= 1.15.1)
  • OpenTelemetry.Instrumentation.Runtime (>= 1.15.1)
  • RabbitMQ.Client.OpenTelemetry (>= 1.0.0-rc.2)
  • Serilog.Extensions.Hosting (>= 10.0.0)
  • Serilog.Formatting.Compact (>= 3.0.0)
  • Serilog.Settings.Configuration (>= 10.0.0)
  • Serilog.Sinks.Async (>= 2.1.0)
  • Serilog.Sinks.Console (>= 6.1.1)
  • Serilog.Sinks.File (>= 7.0.0)
  • Serilog.Sinks.OpenTelemetry (>= 4.2.0)
  • Stratara.Shared (>= 3.2.3)

Three correctness fixes, all of the same kind: behaviour that failed by succeeding. A pipeline
behaviour registered twice ran twice and nothing said so; a closed generic type name normalized to a
malformed key that two different types could share, and the loser of that collision was discarded
silently; and a heavy command republished from the outbox left its lane whenever its type could not
be resolved. None produced an error, and no test covered any of them. They were found by reading the
code against the specification rather than by a failure report.

### Fixed

- **A pipeline behaviour registered twice now installs once.** Every behaviour registrar used a
 plain scoped registration, so a host that called one twice — the ordinary result of composing two
 service bundles that each set up their own slice of the framework — installed the stage twice and
 ran it twice per request. None of the three consequences was benign: every validator ran twice
 against every request, so a validator with a side effect (a uniqueness check against the database)
 performed it twice; the tenant guard ran twice, the second time against a second options instance
 that could disagree with the first; and two nested resilience pipelines multiplied rather than
 added, turning a configured budget of four attempts into sixteen. `AddStrataraValidation`,
 `AddStrataraTenantIsolation`, `AddStrataraResilienceBehavior`, `AddCommandAuditing` and the
 `AddPipelineBehavior` / `AddPipelineBehaviorWithResult` primitives are now idempotent per behaviour
 type. Registering two *different* behaviours is unaffected.
- **A closed generic type name now normalizes correctly.** Resolving a recorded type name is meant to
 ignore the assembly version and match on the type name and assembly name alone. For a closed
 generic it did neither: the name was truncated at its second comma, which for a generic falls
 *inside* the type-argument brackets, so the outer assembly name was dropped and the key became a
 malformed fragment. Two closed generics that differed only in the assembly that declared them
 collapsed onto one key, and the second registration was then silently discarded — an event
 upcaster could match the wrong source type, and a type a host believed it had registered was
 unresolvable. Names are now parsed rather than counted, and each type argument is reduced the same
 way as the outer name, so upgrading the *payload's* assembly no longer strands rows either.
- **A heavy command republished from the outbox stays in the heavy lane.** When a stored command's
 recorded type could not be resolved in the process draining the outbox, republication fell back to
 the shared command topic regardless of the lane the command belonged to. Depending on how the lanes
 are deployed, the command was then either dead-lettered by the interactive worker or executed on
 the interactive lane — the starvation the separate lane exists to prevent. The lane is now recorded
 on the envelope when the command is enqueued, so republication no longer depends on resolving the
 type.

### Changed

- **Registering two different types under one recorded name now fails.** The trusted-type resolver
 previously kept the first and discarded the second without a word, which is indistinguishable from
 a registration that never happened: the type simply fails to resolve later, when a stored row is
 read. `ITrustedTypeResolver.Register` now throws, naming both types. Registering the same type
 again remains a no-op. Reaching this requires two distinct types sharing a full type name *and* a
 simple assembly name in one process.
- **`TenantIsolationOptions` follows the options pattern.** The mode can now be bound from
 configuration with `Configure<TenantIsolationOptions>(section)` in addition to the
 `AddStrataraTenantIsolation(o => ...)` callback, and a second call to the registrar no longer
 leaves a conflicting second options instance behind. `Stratara.Mediator` therefore takes a
 dependency on `Microsoft.Extensions.Options`.

### Added

- **`CommandEnvelope.Heavy`** records whether a command declared itself long-running when it was
 enqueued. Optional and defaulting to `false`, which is how an envelope written by an earlier
 version deserializes; the signed canonical form is unchanged, so existing signatures still verify.