Generate DecompilerSettings boilerplate from [DecompilerSetting] attributes

[siegfriedpammer]

Every version-gated decompiler setting was bookkept in four places that had to stay in sync by hand: the property boilerplate, SetLanguageVersion, GetMinimumRequiredVersion, and the [Category] display string. A survey of the file showed that sync had already drifted in four settings.

This PR adds a second source generator to ICSharpCode.Decompiler.Generators that derives all four from a single [DecompilerSetting] attribute on a partial property: the backing field, the accessors with change notification, the version-derived [Category], and both version methods. DecompilerSettings.cs shrinks from ~2450 to ~990 lines, and adding a new setting is now ~5 lines in one place:

/// <summary>
/// Use C# 9 <c>nint</c>/<c>nuint</c> types.
/// </summary>
[Description("DecompilerSettings.NativeIntegers")]
[DecompilerSetting(CSharp.LanguageVersion.CSharp9_0)]
public partial bool NativeIntegers { get; set; }

[Description] stays handwritten: 51 of the 121 resource keys do not follow the DecompilerSettings.{PropertyName} convention, and the keys are what you grep for from the resx. Settings without a version argument (VB/F#/UI options) keep their handwritten [Category]. New diagnostics keep the scheme honest: DSTG002 (attribute target must be a partial instance bool property), DSTG003 (version-gated settings must not also declare a handwritten [Category], which would make GetCustomAttribute<CategoryAttribute>() throw), DSTG004 (a new LanguageVersion needs a display-category mapping in the generator), DSTG005 (the containing class must be a non-nested partial class).

The first commit is a strict 1:1 translation - the pre-existing inconsistencies are reproduced exactly. The three follow-up commits then remove them, each independently droppable:

1. ExtensionMethods, UseLambdaSyntax and UseEnhancedUsing were disabled by SetLanguageVersion for old targets but never raised GetMinimumRequiredVersion while enabled, unlike every comparable syntax-preference setting.
2. SwitchOnReadOnlySpanChar carried the C# 11.0 category but was missing from both version methods, so decompiling for an older target could still produce switches over ReadOnlySpan<char> that the requested compiler cannot compile.
3. Four settings used the bare category string "Other" instead of the DecompilerSettings.Other resource key; both currently resolve to the same English text, but the keys would split into two options groups the moment their translations diverge.

Verification: a throwaway characterization harness recorded, on the old code, the full property vector after SetLanguageVersion for every LanguageVersion and the GetMinimumRequiredVersion result for each setting toggled alone, and was re-run green against the generated code (red/green around each behavior-changing commit); a reflection diff of the per-property Category/Description/Browsable surface between the old and new build came back identical for all 122 bool properties (also proving the merged partial-property attributes stay unambiguous for the options UI and settings XML round-trip). Full decompiler suite (3365 tests) and ILSpy UI suite (1178 tests) pass.

Later commits from the review rounds re-validate an explicitly set LanguageVersion when project export starts, document the two roles the language version plays in the settings API, and derive the C# 1.0 [Category] strings from the generator's version map as well (observably identical: CSharp1 is the enum floor, so the new cascade bucket cannot fire and the new minimum-version arm returns what the fallback already did).

Identified but out of scope: with UseLambdaSyntax enabled, statement-bodied anonymous functions still decompile to delegate syntax; extending the setting to emit statement lambdas is a candidate follow-up.

---

This PR description was written by the Claude Code agent session that authored the change.

🤖 Generated with Claude Code

[christophwille]

Code review (high effort, multi-agent)

The review of the DecompilerSettings source-generator migration surfaced two real generator robustness bugs: unguarded null-forgiving casts on error-typed attribute arguments crash the entire generator (flooding the build with cascading partial-property errors instead of one targeted diagnostic), and the AddSource hint name drops the namespace half of the (Namespace, ClassName) grouping key, so same-named settings classes in different namespaces would collide deterministically.

The PR also folds two disclosed but untested behavior changes into the refactor: SwitchOnReadOnlySpanChar is newly gated on C# 11, and ExtensionMethods/UseLambdaSyntax/UseEnhancedUsing now raise GetMinimumRequiredVersion -- both of which can change decompiled output or make WholeProjectDecompiler.LanguageVersion assignment throw where it previously succeeded. The 333-line generator with three diagnostics also ships with zero tests.

Remaining findings are cleanups: DSTG003 double-reports via a redundant CS0579, a duplicated EmbeddedAttribute post-init declaration, alias-ambiguous enum-name recovery in diagnostics, and a catch-all diagnostic-id switch that would misreport future ids.

9 findings posted as inline comments (4 correctness, 1 test coverage, 4 cleanup). 16 candidates were independently verified; 15 survived, merged into these 9.

Assisted-by: Claude:claude-fable-5:Claude Code

[christophwille]

Re-review (high effort, multi-agent)

Follow-up to the previous review, against head f7e6a8a.

Six of the nine previous findings are fixed at this head: error-typed attribute arguments are now guarded (skip instead of crash), the hint name carries the namespace, DSTG003 suppresses the generated [Category], the EmbeddedAttribute source is shared via RoslynHelpers, the enum name is recovered from syntax (alias-safe), and the diagnostic-id lookup is a fail-loud dictionary. Nice turnaround.

Three findings carry over: the two behavior changes (ExtensionMethods/UseLambdaSyntax/UseEnhancedUsing raising the minimum version; SwitchOnReadOnlySpanChar gated on C# 11) are now explicit, well-described commits -- but both still ship without a regression test pinning the new semantics, and the generator itself remains untested.

Five new findings: two latent generator input-validation gaps (invalid output for global-namespace classes and for camelCase-named settings), a missing containing-type check, duplicated LanguageVersion validation in WholeProjectDecompiler, and 14 handwritten C# 1.0 [Category] literals the generator's category map could single-source.

8 findings posted as inline comments (4 correctness, 1 test coverage, 3 cleanup). 18 candidates were independently verified; the one candidate targeting the since-removed AffectsMinimumRequiredVersion knob was dropped as obsolete (the branch moved mid-review).

Assisted-by: Claude:claude-fable-5:Claude Code