0.4: modernize the API surface for current .NET conventions

[Agash]

Tracking the 0.4 work. Breaking changes are acceptable as long as callers update lines, not rewrite.

Landed

API shape

• The whole surface is grouped by protocol category: client.Scenes.GetSceneListAsync(...), with conveniences sitting in the same group as the requests they wrap. The flat extension methods are gone.
• Typed batch builder over the same categories, returning BatchRef<T>; results read with results.Get(reference)
• Per-event IAsyncEnumerable streams alongside the classic events
• Real C# enums for string-valued protocol enums; constants to const
• Non-nullable response types (CallRequiredAsync)
• Typed exceptions carrying the OBS status
• WaitForEventAsync throws TimeoutException; overloads for waiting without a predicate
• SwitchProgramSceneAsync / SwitchPreviewSceneAsync replace a bool parameter
• TryGetSceneItemIdAsync to FindSceneItemIdAsync; int overloads for scene item ids
• C# 14 extension members

Host and diagnostics

• AddObsWebSocketClient(name) reading ConnectionStrings
• WithAutoConnect() hosted service; an unreachable OBS at startup is survivable
• AddHealthChecks().AddObsWebSocket()
• Configuration changes apply without a restart, via IOptionsMonitor
• Keyed DI for multiple OBS instances
• Options validated on resolve, AOT-safe
• ActivitySource, and metric instruments built from IMeterFactory
• [LoggerMessage] throughout with typed parameters, no suppression
• TimeProvider for every timeout and reconnect delay
• Reconnect backoff via Microsoft.Extensions.Resilience, with jitter

Fixes found along the way

• SetSceneItemEnabled never sent sceneName (positional binding after the 5.7 refresh)
• CustomEvent payload never deserialized
• Batch results mis-paired on MessagePack; each result is now parsed in isolation
• Process crash from a timeout timer firing after its source was disposed
• Fields the protocol documents as nullable in prose were generated non-nullable, so GetMediaInputStatus failed to deserialize whenever media was not playing
• EventStream deferred argument validation

Decided against

• Live state projection. OBS is not reliable enough as a source of truth to cache; read at the point of use.
• IObsWebSocketClient. A source-generated sealed client does not need one, and a narrow interface would mock the low-level calls while leaving the helpers unmockable. Test at the transport instead (Testing package with an in-memory OBS server #18).
• Shape-matching to recover parallel batch results. 75 of 147 requests return no payload, five groups share identical response shapes, and requestStatus is displaced too. See OBS mis-pairs batch results under Parallel execution (upstream bug) #16.
• params ReadOnlySpan on the raw batch overload. params must be last, so it cannot coexist with executionType and cancellationToken.

Deferred

• OBS mis-pairs batch results under Parallel execution (upstream bug) #16 parallel batch mis-pairing, upstream in obs-websocket
• Ship a Roslyn analyzer with the package #17 Roslyn analyzer shipped with the package
• Testing package with an in-memory OBS server #18 testing package with an in-memory OBS server
• Handles for addressing scenes, inputs, filters and scene items #19 handles for addressing scenes, inputs, filters and scene items
• Timing-sensitive tests are flaky under parallel workers; they pass individually and on CI

[Agash]

Auditing the built assembly rather than the source turned up items above that were checked but not actually finished, plus work that is not listed. #20 covers it.

Boxes that were not true when ticked

• "The flat extension methods are gone" was half true. 60 per-event stream accessors were still flat extensions on the client, so requests and conveniences were grouped and events were not. They now sit on their category group, generated from the same category field in protocol.json that the requests use. {Category}RequestGroup is {Category}Group, since the groups no longer hold only requests.
• "SwitchProgramSceneAsync / SwitchPreviewSceneAsync replace a bool parameter" added them but left SwitchSceneAsync(..., bool switchToProgram) public next to them, so the group appeared to offer five ways to switch a scene. The flag versions are private now, and the missing SwitchPreviewSceneAndWaitAsync exists.
• "int overloads for scene item ids" was the wrong shape. Passing an int already worked everywhere without overloads, because int widens to double implicitly. Reading was the actual problem, and the overload only fixed one helper while GetSceneItemIdAsync().SceneItemId still returned double. See below.

Not previously listed

• Whole-number protocol fields are int or long rather than double. Which fields those are cannot be recovered from the definition, since sceneItemId and inputVolumeMul are both Number with a >= 0 restriction, so the classification is an explicit table rather than a rule over field names. 32 field names retyped, 10 deliberately left fractional, and an unlisted Number field reports OBSWSGEN012 instead of drifting in.
• AddObsWebSocketClient returns a builder, so WithAutoConnect / WithHealthCheck / WithReconnectPipeline chain off it. The old WithAutoConnect hung off IServiceCollection and could not target one of two named clients.
• The RequestStatus enum is RequestStatusCode, because it shared a name with the RequestStatus record carried on every response, which is exactly the pair a caller has to disambiguate to write a status filter.
• Six helpers decided "not found" by substring matching the exception message. ObsWebSocketRequestException.StatusCode exposes the reported status as the protocol enum and they match on that.
• SetInputMutesAsync still built raw BatchRequestItems and returned nothing, so per-input failures were log-only. It uses the typed builder and returns BatchResults.
• csharpier was pinned in dotnet-tools.json and never run. 69 of 99 files had drifted, some of it misleading rather than cosmetic. CI checks it now.

Fixes found along the way

• GetSourceScreenshot returns a data URI, not bare Base64, so GetSourceScreenshotBytesAsync threw and returned null on every call since it was written. The file variant was unaffected, which is why it went unnoticed. Found only by asserting the returned bytes rather than that the call succeeded.

Deferred item that is now done

• "Timing-sensitive tests are flaky under parallel workers; they pass individually and on CI" understated it. Three had real defects: two drove a FakeTimeProvider and then waited on a real Task.Delay, which defeats the fake clock and cannot finish inside the test's own timeout under load, and one put a 1s wall clock limit on a test that only builds a container. Fixed at source. A residual set still times out when all three target frameworks run at once, which looks like thread pool starvation from running ProcessorCount workers per framework rather than a defect in any one test.