feat: Add inline suppression for LCOM4 class cohesion

[fummicc1]

Summary

Extends // swift-complexity:disable to type declarations (class/struct/actor) so a reviewed low-cohesion type can be exempted from LCOM4 judgments. Previously the only workaround was --exclude, which hides the whole file's complexity violations too.

Closes #36.

Syntax

// swift-complexity:disable lcom4
class LegacyOrderManager { ... }  // member functions are still checked

// swift-complexity:disable
struct ScatteredHelper { ... }    // bare form: suppresses the type's own metrics (= lcom4)

Design

• Scope intersection: SuppressionParser now takes an applicableTo scope and intersects parsed metrics with it — FunctionDetector passes functionLevel (cyclomatic/cognitive), NominalTypeDetector passes typeLevel (lcom4). This yields three properties at once:
  • A bare disable means "every metric of this declaration itself": above a type it suppresses only lcom4 and never cascades to member functions (a dedicated test pins this).
  • A metric of the wrong level (lcom4 above a function, cyclomatic above a class) suppresses nothing — the fail-closed contract from feat: Add inline suppression comments for complexity thresholds #35 extends naturally.
  • JSON backward compatibility: without the intersection, adding lcom4 to SuppressedMetric.allCases would have leaked "lcom4" into the suppressedMetrics array of bare-disabled functions, changing v1.1.0 output. A test now pins the function set to {cyclomatic, cognitive}.
• Values stay visible, judgments are skipped (same contract as feat: Add inline suppression comments for complexity thresholds #35): ClassCohesion gains an optional suppressedMetrics set; the LCOM4 value is still computed and shown everywhere. All three cohesion judgment sites are gated — the Xcode diagnostic, SARIF's lcom4_cohesion results, and the CLI's threshold filtering. The exit code needs no change because LCOM4 never influenced it.
• --report-suppressions now also lists suppressed types with current LCOM4 values, and its help text mentions types.
• CohesionSummary statistics and cohesionLevel classification are deliberately unchanged — they are data, not judgments.

Verification

• swift test: 107 tests pass, including the first direct unit tests for NominalTypeDetector (type-level scoping fixture: bare/lcom4-scoped/unsuppressed/wrong-level/typo cases), a no-cascade test over member functions, parser scope-intersection tests, and a formatter test asserting a suppressed LCOM4=5 class produces no SARIF result and no Xcode diagnostic while remaining visible in JSON
• E2E against the real LCOM4 path (--lcom4 --index-store-path): a suppressed LCOM4=3 class disappeared from xcode/sarif output while an identical unsuppressed twin still reported; --report-suppressions listed it; JSON kept both values
• Function-level JSON output verified unchanged against the v1.1.0 fixtures (no lcom4 leakage)
• swift-format clean; markdownlint error count identical to main

https://claude.ai/code/session_01GaTHVDhSAx76RbLmVziLQu

[github-actions]

📊 PR Analysis

Type	Count
Swift files	12
Test files	3
Documentation	2
Configuration	0

📝 Large PR: Consider breaking this into smaller, focused PRs.

---

Automated analysis by swift-complexity CI

[github-actions]

🔍 Complexity Analysis Results

✅ All functions are below the complexity threshold of 15.

This analysis was performed by swift-complexity itself.