fix(LC0089): keep compilation-scoped settings out of analyzer instance fields

[Arthurvdv]

Problem

Issue #254 audited the RegisterCodeBlockAction analyzers for incremental-compilation staleness. Verification against the decompiled NAV SDK source (18.0.x) refuted the audit's core premise — the declaration-cache skipping machinery is dead code (null AnalysisState on every shipping driver path; CompilationWithAnalyzers never instantiated), there is no SyntaxNodeAction-vs-CodeBlockAction asymmetry, and the "LC0059 settings staleness" recommendation is moot because ALCopsSettingsProvider never invalidates its cache anyway. Full analysis: issue comment.

The verification did surface one real defect: CognitiveComplexity (LC0089/LC0089i/LC0090) stores per-compilation state in mutable instance fields, written at CompilationStart. Analyzer instances are materialized once per project and shared across analysis passes, so an overlapping pass or sibling project with a different alcops.json / ruleset can overwrite complexityThreshold and the LC0089i enablement flag mid-analysis — the same bug class as the LC0054 fix, and a violation of the repo's stateless-analyzer rule.

Solution

• Capture both values as locals in the CompilationStart closure (like the existing recursion graph service) and thread them as parameters; the affected methods became static. No functional change under single-pass analysis.
• Rewrite the incremental-compilation section of .claude/rules/sdk-analyzer-infrastructure.md to the verified model: no per-declaration cache skipping exists; the real callback filters are the per-file AnalysisScope and the module-only partial pass, uniform across all action kinds — which is also the actual mechanism behind the [Bug]: AC0032 - false positive #243 AC0032 false positives (the accumulator was scope-starved, not cache-skipped; the fix(AC0032): rewrite to atomic per-object analysis with variable-map optimization #253 fix remains correct, with corrected rationale). The two-phase accumulator anti-pattern verdict stands.
• New rule doc .claude/rules/diagnostics/lc0089-cognitive-complexity.md recording the design decisions.

Testing

• No new fixture: the race requires overlapping compilation passes against different settings, which unit tests cannot reproduce; existing tests pin the functional behavior (14/14 CognitiveComplexity tests pass).
• Full solution: 1,593 passed, 0 failed.
• Three-TFM Release build of ALCops.LinterCop (ContinuousIntegrationBuild=true): clean.
• dotnet format --verify-no-changes: no findings.

Closes #254

🤖 Generated with Claude Code