Fix StackOverflow crash on cyclic local initializers (MA0091/MA0092/MA0093)

[meziantou]

Fixes #1323

What changed

EventsShouldHaveProperArgumentsAnalyzer.EventReferenceVisitor.FindFromLocalSymbol resolved a local's initializer and, when that initializer bound to another ILocalSymbol, called itself — with no visited set. Two locals whose initializers reference each other made the walk recurse until the process died:

class C
{
    void M()
    {
        System.EventHandler a = b;
        System.EventHandler b = a;
        a.Invoke(this, System.EventArgs.Empty);
    }
}

FindFromLocalSymbol now tracks the visited locals in a HashSet<ISymbol> (with SymbolEqualityComparer.Default) held by the visitor instance and bails on re-entry.

The two hardcoded CancellationToken.None arguments are also gone: the visitor now takes the real context.CancellationToken through a primary constructor, so the token parameter disappears from the recursive method's signature. With a cycle present, cancellation was the only thing that could have stopped the walk.

Why it matters

A StackOverflowException cannot be caught by Roslyn's analyzer host, so this was not a recoverable AD0001 — the whole process died (VBCSCompiler, Visual Studio, Rider, OmniSharp). MA0091, MA0092 and MA0093 are all enabled by default at Warning.

The source that triggers it does not compile (CS0841, use before declaration), which is exactly the state of a file while it is being typed — and IDE live analysis runs continuously on erroneous compilations.

Test infrastructure

ProjectBuilder.IsValidCode, set by WithNoCompilation(), was declared but never read: GetSortedDiagnostics hardcoded compileSolution: true, so no analyzer test could use non-compiling source (Assert.Fail("The code doesn't compile.")). It now passes IsValidCode, which is what makes the regression tests possible. No existing test used WithNoCompilation(), so nothing else changes behavior.

Tests

Two regression tests: the mutual a = b; b = a; cycle from the issue, and a self-referencing EventHandler a = a;.

I confirmed the new test actually reproduces the bug — with the visited-set guard temporarily removed, the run died exactly as reported in the issue (Test run summary: Zero tests ran, exit code 7, createdump). Guard restored, it passes.

• 12/12 tests in EventsShouldHaveProperArgumentsAnalyzerTests pass on all five Roslyn versions (4.8, 4.14, 5.0, 5.6, 5.9)
• Full suite on the default version: 3776/3776 passed
• dotnet run --project src/DocumentationGenerator produced no markdown changes