AspectGenerator: Instructions file for Codex

Agent Instructions

Before working on AspectGenerator hardening, diagnostics, code generation, or incremental generator refactors, read docs/AspectGeneratorHardeningPlan.md.

Current notes:

• Invocation discovery should stay on SyntaxProvider candidates. Do not reintroduce full syntax-tree scans through CompilationProvider + DescendantNodes().
• Keep InterceptMethods diagnostics deferred unless the user explicitly asks to work on them.
• Preserve CRLF line endings for touched repository files.
• Do not add new NuGet packages, template engines, formatters, source-generation helper libraries, or build tools unless the maintainer explicitly asks for that dependency.
• Do not replace internal code generation with Scriban, Handlebars.Net, CSharpier, or other external templating/formatting tools as an automatic refactoring step.

Target filter parser

The parser of the target filter language is written in DotGram (Source/TargetFilter/TargetFilter.gram) and generated ahead of the build. The generated C# in Source/TargetFilter/Generated is committed and compiled like any other source: the build and the package have no DotGram dependency, and nothing is generated at build time.

Every target filter is read by the generated parser (Source/TargetFilter/TargetFilterFront.cs compiles what it reads into the rules the matchers run). The golden baselines in Baselines/TargetFilters pin what it reads over the whole corpus (UnitTests/Tests/TargetFilters): the diagnostic ids, the plan, the compiled rules and the decision for every target of a fixed list. The generated parser compiles against Roslyn 4.12 on netstandard2.0, and the minimum-sdk consumers run it in the SDK 9.0.100 compiler through their TargetFilter.

• The grammar reads a line in the stages the filter language has always read it: a line, then each piece of it the language trims and reads again (a condition item, a value, a dotted name, the words of a method pattern, a parameter list and its parameters), each a publication read over the trimmed piece. Keep that shape when changing the language, and run the golden tests: a baseline difference is either a defect of the change or an intended change that needs a changelog line and re-accepted baselines (AG_UPDATE_BASELINES=1), with every changed record explained in the commit.
• The generated parser can carry a reading nested deeper than its stack over to a thread of its own; TargetFilterParser.OnHandingOff refuses that (the generator runs on the compiler's threads), so such a rule is reported as AG0202. That code reads the thread's cultures, which RS1035 bans in analyzers; Source/TargetFilter/.editorconfig silences RS1035 for the generated file alone, until DotGram stops reading them there.