release: 1.4.0

CHANGELOG.md

@@ -5,104 +5,9 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [1.4.0-rc.2] — 2026-09-16
+## [1.4.0] — 2026-10-13
 
-> Companions: `OpenTelemetry` 1.1.1-rc.2 · `HybridCache` 1.0.10-rc.2 · `FluentValidation` 1.0.10-rc.2.
-
-No behavior changes beyond rc.1. Ten defects found by the Pipeline Explorer team, who measured each
-one against packages built from rc.1 and sent a reproduction with every report.
-
-All five packages move together even though three of them changed only their README. The companions
-reference the core by project, so a published companion pins the core version it was packed against:
-leaving `HybridCache` at rc.1 would keep pulling the rc.1 core — with the generator defect below —
-into any application that installs the companion without naming the core itself.
-
-### Added
-
-- **`MediatorCapabilities`** — a public declaration of what a build of the core guarantees, for source
-  generators written against it. `ChainLifetimeFold` is the first: the lifetime of a handler constrains
-  the lifetime of the pipeline chain that consumes it.
-
-  Tooling that registers its own pipeline components has to know this to pick their lifetime, and until
-  now it could only infer it — one generator was reading the presence of `IMediatorDispatchObserver` as
-  evidence of the fold, because the type and the fix shipped in the same commit. That inference would
-  have survived a rename by luck and failed silently otherwise, and the failure lands in the
-  consumer's `BuildServiceProvider` rather than at build time.
-
-  It lives in the core rather than in `Abstractions` on purpose. The core declares a dependency on a
-  MINIMUM version of `Abstractions`, not a pinned one, so NuGet can resolve an old core beside a new
-  `Abstractions` — and a capability declared there would then describe a core that does not have it.
-  Constants are `int` rather than `bool` so a capability can advance without needing a new name, and
-  `const` rather than `static readonly` so a generator can read the value through
-  `IFieldSymbol.ConstantValue`.
-
-
-### Fixed
-- **Interceptors could rewrite a call inside an expression tree when the type name was ambiguous.**
-  The guard that keeps `Send`, `Publish` and `CreateStream` out of expression trees — the reason a Moq
-  `Setup` or `Verify` still works — resolved `Expression<T>` with `GetTypeByMetadataName`. That returns
-  null both when the type is absent and when it is declared in more than one referenced assembly, and
-  the two are not the same: with the type absent a consumer cannot write an expression tree at all,
-  but with it ambiguous they can, and the guard opened. Reachable by a project pulling the legacy
-  `System.Linq.Expressions` package alongside the framework one. Now resolved with the plural
-  `GetTypesByMetadataName`, matching against any candidate.
-
-
-- **A stream behavior that narrows its own constraint broke the consumer's build.** `StreamGenerator`
-  closed it over every pair, producing `CS0311` in `StreamRegistry.g.cs`, a file nobody can edit. The
-  narrowing data had been computed for stream behaviors all along — the scanner does not branch on the
-  interface kind — and this generator never read it. The equivalent guard on the request side shipped
-  in rc.1; this is the half that was missing.
-- **A `file`-local stream behavior reached the generated registry**, through a hand-rolled
-  accessibility test: a `file` type reports `Internal`, so "Public or Internal" waved it through and
-  the generated file named something the compiler cannot resolve — `CS0400`. It now uses the shared
-  check, which also covers private nested types and types nested in a generic one.
-- **DSOFT011 described a consequence the developer could not reach.** For that same stream case it
-  reported the behavior and said it "works on the ordinary runtime but throws under Native AOT", while
-  the same generator run emitted code that does not compile. Fixing the generator makes the message
-  true again.
-- **A stream whose consumer stopped reading was reported as failed.** `break` after the first page
-  disposed the iterator and the span went red — a paged query reading its first N rows looked like an
-  outage.
-- **No failing stream carried `error.type` or an exception event**, and cancelling, breaking and
-  crashing produced spans identical character for character. Streams now report
-  `mediator.stream.termination` as `completed`, `cancelled`, `early` or `faulted`, and only a fault
-  sets Error status.
-- **An empty stream reported `first_item_ms` and `throughput_per_sec` as zero**, indistinguishable
-  from measurements and averaged as measurements. Both tags are now omitted when nothing was produced.
-- **`mediator.request.errors` was never incremented for a stream.** Error rate stayed flat while the
-  traces beside it showed failures. C# forbids a catch clause in an iterator containing `yield return`,
-  so both stream behaviors now drive the enumerator by hand — awaiting `MoveNextAsync` inside its own
-  try and yielding outside every try, which is the shape that makes a catch legal.
-- **Spans opened inside a stream handler escaped the stream span from the second item onward.**
-  `Activity.Current` is an `AsyncLocal`, and after the first `yield return` the iterator resumes under
-  the consumer's context. A stream created under one span and enumerated after it ended came out as a
-  root span in a fresh trace — the ordinary shape of any API handing an `IAsyncEnumerable` to a
-  framework that enumerates it later. The span is now made ambient around every `MoveNextAsync`.
-- **The notification error metric lost `mediator.request.type`.** It read the value back off the
-  envelope span, which returns null whenever that span carries no tags — not only when there is no
-  span, since tags are set under `IsAllDataRequested` and any head sampler returning `PropagationData`
-  leaves the envelope alive and untagged.
-- **The README's manual registration armed one dispatch table of three.** `AddMediator()
-  .RegisterMediatorHandlers().PrecompilePipelines()`. The two you can forget fail very differently:
-  can forget fail very differently. `CreateStream` throws and names the missing call. `Publish` says
-  nothing at all — it reaches no handlers and returns a completed task, because an empty table is
-  indistinguishable from an application that has no notifications. The generated
-  `AddMediator(configure)` overload calls all three.
-
-### Documentation
-
-- The four packaged READMEs, which are the nuget.org pages. The OpenTelemetry one described
-  instrumentation as attaching "via standard pipeline behaviors", which rc.1 replaced: `Send` and
-  `Publish` are observed through the core's observation ports and only stream tracing and metrics are
-  behaviors. It also gained the registration-order section the other companions have. The HybridCache
-  one gained the Native AOT section it never had — it is the one companion that fails there without a
-  step from the application, and DSOFT012 shipped in rc.1 unmentioned.
-
-
-## [1.4.0-rc.1] — 2026-09-02
-
-> Companions: `OpenTelemetry` 1.1.1-rc.1 · `HybridCache` 1.0.10-rc.1 · `FluentValidation` 1.0.10-rc.1.
+> Companions: `OpenTelemetry` 1.1.1 · `HybridCache` 1.0.10 · `FluentValidation` 1.0.10.
 
 **Four** changes in this release alter observable behavior. Each is marked *This is a behavior
 change* under **Changed**, and all four are worth reading before you upgrade:
@@ -184,6 +89,23 @@ change* under **Changed**, and all four are worth reading before you upgrade:
   scaling suite, `run-all-benchmarks.cmd`, and `compare-runs.ps1`, which compares two runs by overhead
   above each suite's own baseline so machine drift cancels out.
 
+- **`MediatorCapabilities`** — a public declaration of what a build of the core guarantees, for source
+  generators written against it. `ChainLifetimeFold` is the first: the lifetime of a handler constrains
+  the lifetime of the pipeline chain that consumes it.
+
+  Tooling that registers its own pipeline components has to know this to pick their lifetime, and until
+  now it could only infer it — one generator was reading the presence of `IMediatorDispatchObserver` as
+  evidence of the fold, because the type and the fix shipped in the same commit. That inference would
+  have survived a rename by luck and failed silently otherwise, and the failure lands in the
+  consumer's `BuildServiceProvider` rather than at build time.
+
+  It lives in the core rather than in `Abstractions` on purpose. The core declares a dependency on a
+  MINIMUM version of `Abstractions`, not a pinned one, so NuGet can resolve an old core beside a new
+  `Abstractions` — and a capability declared there would then describe a core that does not have it.
+  Constants are `int` rather than `bool` so a capability can advance without needing a new name, and
+  `const` rather than `static readonly` so a generator can read the value through
+  `IFieldSymbol.ConstantValue`.
+
 ### Changed
 
 - **Pipeline components registered through `MediatorBuilder` now default to `Scoped`, not
@@ -360,12 +282,72 @@ change* under **Changed**, and all four are worth reading before you upgrade:
   default (no publisher is registered by default), and `PrecompilePipelines()` was described as
   capturing the component list (it records only whether a chain is needed, and freezes its lifetime).
 
+- **Interceptors could rewrite a call inside an expression tree when the type name was ambiguous.**
+  The guard that keeps `Send`, `Publish` and `CreateStream` out of expression trees — the reason a Moq
+  `Setup` or `Verify` still works — resolved `Expression<T>` with `GetTypeByMetadataName`. That returns
+  null both when the type is absent and when it is declared in more than one referenced assembly, and
+  the two are not the same: with the type absent a consumer cannot write an expression tree at all,
+  but with it ambiguous they can, and the guard opened. Reachable by a project pulling the legacy
+  `System.Linq.Expressions` package alongside the framework one. Now resolved with the plural
+  `GetTypesByMetadataName`, matching against any candidate.
+
+
+- **A stream behavior that narrows its own constraint broke the consumer's build.** `StreamGenerator`
+  closed it over every pair, producing `CS0311` in `StreamRegistry.g.cs`, a file nobody can edit. The
+  narrowing data had been computed for stream behaviors all along — the scanner does not branch on the
+  interface kind — and this generator never read it. The request side gained the same guard in this
+  release; both halves land together.
+- **A `file`-local stream behavior reached the generated registry**, through a hand-rolled
+  accessibility test: a `file` type reports `Internal`, so "Public or Internal" waved it through and
+  the generated file named something the compiler cannot resolve — `CS0400`. It now uses the shared
+  check, which also covers private nested types and types nested in a generic one.
+- **A stream whose consumer stopped reading was reported as failed.** `break` after the first page
+  disposed the iterator and the span went red — a paged query reading its first N rows looked like an
+  outage.
+- **No failing stream carried `error.type` or an exception event**, and cancelling, breaking and
+  crashing produced spans identical character for character. Streams now report
+  `mediator.stream.termination` as `completed`, `cancelled`, `early` or `faulted`, and only a fault
+  sets Error status.
+- **An empty stream reported `first_item_ms` and `throughput_per_sec` as zero**, indistinguishable
+  from measurements and averaged as measurements. Both tags are now omitted when nothing was produced.
+- **`mediator.request.errors` was never incremented for a stream.** Error rate stayed flat while the
+  traces beside it showed failures. C# forbids a catch clause in an iterator containing `yield return`,
+  so both stream behaviors now drive the enumerator by hand — awaiting `MoveNextAsync` inside its own
+  try and yielding outside every try, which is the shape that makes a catch legal.
+- **Spans opened inside a stream handler escaped the stream span from the second item onward.**
+  `Activity.Current` is an `AsyncLocal`, and after the first `yield return` the iterator resumes under
+  the consumer's context. A stream created under one span and enumerated after it ended came out as a
+  root span in a fresh trace — the ordinary shape of any API handing an `IAsyncEnumerable` to a
+  framework that enumerates it later. The span is now made ambient around every `MoveNextAsync`.
+- **The README's manual registration armed one dispatch table of three.** `AddMediator()
+  .RegisterMediatorHandlers().PrecompilePipelines()`. The two you can forget fail very differently:
+  can forget fail very differently. `CreateStream` throws and names the missing call. `Publish` says
+  nothing at all — it reaches no handlers and returns a completed task, because an empty table is
+  indistinguishable from an application that has no notifications. The generated
+  `AddMediator(configure)` overload calls all three.
+
 ### Removed
 
 - **`BehaviorLinkRegistry`** — a middle chain-construction tier no generator ever populated. Its
   dictionary was null in every process and `Link()` fell through to the interface-typed adapter on
   every call. Per-pair generic static types: 6 → 5.
 
+### Documentation
+
+- **The four packaged READMEs**, which are the pages a reader sees on nuget.org, were checked against
+  the code rather than against the docs site.
+
+  The OpenTelemetry one described instrumentation as attaching "via standard pipeline behaviors". That
+  is no longer how it attaches: `Send` and `Publish` are observed through the core's observation ports
+  and only stream tracing and metrics are behaviors. The difference is not cosmetic — the request span
+  is opened at the dispatch boundary, so it wraps pre- and post-processors, which a behavior cannot
+  reach. It also gained the registration-order section the other companions already had.
+
+  The HybridCache one gained the Native AOT section it never had. It is the one companion that fails
+  there without a step from the application, and DSOFT012 exists to say so at build time.
+
+  `Abstractions` was the only one of the four without the logo header.
+
 ## [1.3.0] — 2026-07-08
 
 > Companions: `OpenTelemetry` 1.1.0 · `HybridCache` 1.0.9 · `FluentValidation` 1.0.9.

src/DSoftStudio.Mediator.Abstractions/DSoftStudio.Mediator.Abstractions.csproj

@@ -12,7 +12,7 @@
 
 	  <!-- NuGet -->
 	  <PackageId>DSoftStudio.Mediator.Abstractions</PackageId>
-	  <Version>1.4.0-rc.2</Version>
+	  <Version>1.4.0</Version>
 
 	  <Authors>DSoftStudio</Authors>
 	  <Company>DSoftStudio</Company>

src/DSoftStudio.Mediator.FluentValidation/DSoftStudio.Mediator.FluentValidation.csproj

@@ -9,7 +9,7 @@
 
 		<!-- NuGet -->
 		<PackageId>DSoftStudio.Mediator.FluentValidation</PackageId>
-		<Version>1.0.10-rc.2</Version>
+		<Version>1.0.10</Version>
 
 		<Authors>DSoftStudio</Authors>
 		<Company>DSoftStudio</Company>

src/DSoftStudio.Mediator.HybridCache/DSoftStudio.Mediator.HybridCache.csproj

@@ -9,7 +9,7 @@
 
 		<!-- NuGet -->
 		<PackageId>DSoftStudio.Mediator.HybridCache</PackageId>
-		<Version>1.0.10-rc.2</Version>
+		<Version>1.0.10</Version>
 
 		<Authors>DSoftStudio</Authors>
 		<Company>DSoftStudio</Company>

src/DSoftStudio.Mediator.OpenTelemetry/DSoftStudio.Mediator.OpenTelemetry.csproj

@@ -9,7 +9,7 @@
 
 		<!-- NuGet -->
 		<PackageId>DSoftStudio.Mediator.OpenTelemetry</PackageId>
-		<Version>1.1.1-rc.2</Version>
+		<Version>1.1.1</Version>
 
 		<Authors>DSoftStudio</Authors>
 		<Company>DSoftStudio</Company>

src/DSoftStudio.Mediator/DSoftStudio.Mediator.csproj

@@ -23,7 +23,7 @@
 
 		<!-- NuGet -->
 		<PackageId>DSoftStudio.Mediator</PackageId>
-		<Version>1.4.0-rc.2</Version>
+		<Version>1.4.0</Version>
 
 		<Authors>DSoftStudio</Authors>
 		<Company>DSoftStudio</Company>