Update roslyn to 5.9.0-1.26308.3

[davidwengier]

View Complete Diff of Changes

• Unsafe evolution: propagate unsafe context to constructor initializer (PR: #83969)
• Parameterize DartLab templates repository alias (PR: #84037)
• Fix frontmatter syntax in Compiler.instructions.md (PR: #84032)
• Add baseline availability status to HTML completion tooltips (PR: #84054)
• Fix Razor breakpoint placement (PR: #84029)
• Fix a number of helix issues (PR: #84039)
• Use xsd:anyURI type to detect external completion instead of vs:preferredextensions (PR: #84033)
• Add formatting tests for code blocks above markup (PR: #84027)
• Move Razor docs under docs (PR: #84024)
• Add test to demonstrate adding using for inject directives works (PR: #84026)
• Fix incremental build in Microsoft.Build.Tasks.CodeAnalysis.UnitTests (PR: #84016)
• Remove unneeded logging (PR: #84023)
• Convert copilot-instructions to AGENTS.md (PR: #84022)
• SourceMemberMethodSymbol.IsMetadataVirtual should force complete declaring type when queried by a different module (PR: #84013)
• Free pooled objects across cancellation exceptions (PR: #83875)
• Bump Debugger.Contracts to 18.3.0-beta.26277.4 (PR: #84018)
• Enable parallel builds in the LSP when loading projects (PR: #83982)
• Fix Razor diagnostics exception when HTML diagnostic range exceeds SourceText bounds (PR: #84002)
• Use MEF services in VS Code cohost tests (PR: #83986)
• Implement runtime async support for dynamic (PR: #83713)
• Remove IDynamicFileInfoProvider and related code (PR: #83928)
• Add IAsyncLifetime overhead adjustment to test scheduling (PR: #83915)
• Move a couple of helpers to the remote project (PR: #83990)
• Unsafe evolution: consolidate reserved attributes (PR: #83908)
• Fix rename in VS Code from a C# document (PR: #83971)
• Remove invalid casts in PairedExtensionOperatorSignatureComparer (PR: #83934)
• Remove the state from MSBuildDiagnosticLogger (PR: #83977)
• Unions: not pattern is applied to the incoming value itself (PR: #83904)
• Complete server shutdown when mutex is closed (PR: #83921)
• Fix some missing attributes on partial async and iterator methods (PR: #83790)
• Unsafe evolution: safe/unsafe externs (PR: #83661)
• Use params ImmutableArray<T> on all applicable public APIs. (PR: #83858)
• Unsafe evolution: avoid unnecessary diagnostics on params collection declarations (PR: #83641)
• Cache Razor cohost MEF setup across tests (PR: #83936)
• Consolidate Linked Editing Range up to its callers (PR: #83956)
• Consolidate Semantic Tokens code up to where it should be (PR: #83955)
• Consolidate Diagnostics code up into Remote.Razor (PR: #83951)
• Consolidate Rename code up into Remote.Razor (PR: #83950)
• Consolidate Document Symbols code up into Remote.Razor (PR: #83949)
• Consolidate Folding Ranges code up into Remote.Razor (PR: #83948)
• Consolidate Spell Check code up into Remote.Razor (PR: #83947)
• Per server workspace and project factories (PR: #83857)
• Fix marketplace name for dotnet/skills (PR: #83961)
• Only add reference counts on CodeLens elements that are supported (PR: #82641)
• [main] Update dependencies from dotnet/arcade (PR: #83941)
• Closed classes: add a few interface matching tests (PR: #83940)
• Fix x86 test host OOM in TemporaryStorageService (PR: #83952)
• Change insertionCreateDraftPR to false (PR: #83958)
• Allow running DartLab as VS test gate (PR: #83690)
• Remove ExternalAccess.Razor and give Razor deeper IVT to Roslyn (PR: #83879)
• Remove catch from FBA fuzz test (PR: #83820)
• Move cohost handlers from ExternalAccess.Razor to Razor (PR: #83880)
• Empty and Remove ExternalAccess.Razor (PR: #83920)

[davidwengier]

All of the C# legs are failing, so looks like a real bug. Copilot had this to say:

I found the likely low-risk mitigation: the only LSP parallelism knob is the ProducerConsumer.RunParallelAsync in LanguageServerProjectLoader; serializing that path would specifically revert the risky behavior while preserving the rest of the build-host refactor.The CI failures look like a real Roslyn regression, but not a completion-provider bug.

The common failure is that the LSP sometimes loads app.csproj without framework metadata references after Roslyn’s recent parallel project-loading changes. In the failing completion case, completion.cs is correctly found in Host; project app (net8.0), but the immediately preceding C#.log project-load entry for app.csproj has source/analyzer files and no metadata reference line, so symbols like System.Console can’t bind. That also explains the other failures: go-to-definition can’t navigate to Console, semantic tokens classify framework identifiers incorrectly, diagnostics/hints lose semantic context, and completion doesn’t offer Console.

Most likely culprit is dotnet/roslyn#83982:

• removed the global ProjectBuildManager semaphore
• enabled parallel LSP project loading via ProducerConsumer.RunParallelAsync
• raised LSP build-host maxNodeCount

The logs show a batch where lib, app, and test are loaded concurrently; one project sometimes gets full references while another gets incomplete command-line/reference data. A low-risk mitigation would be to serialize the LSP project-load loop again, rather than changing completion.

I'm guessing "just serialize project loading" is probably not the nuanced solution we want here. Thoughts @jasonmalinowski or @dibarbet ? Do we need to change the tests on this side to wait for a new/better signal, or is this a real product bug being discovered?

[dibarbet]

I'm guessing "just serialize project loading" is probably not the nuanced solution we want here. Thoughts @jasonmalinowski or @dibarbet ? Do we need to change the tests on this side to wait for a new/better signal, or is this a real product bug being discovered?

I think it could be a product bug, and definitely needs a server side fix. The tests wait on this - https://github.com/dotnet/roslyn/blob/576f18b3e54f0511c2fc93a2b20d2762148b688b/src/LanguageServer/Microsoft.CodeAnalysis.LanguageServer/HostWorkspace/LanguageServerProjectSystem.cs#L94

Possibly the parallel load is doing async work that isn't being tracked, or the parallel load is simply failing to fully load the projects. @jasonmalinowski would you mind taking a look?

[JoeRobich]

This insertion PR is for the build just prior to where the build was parallelized #9404