sync: catch up scoring with pyscn (initial catch-up 2/3)

[DaisukeYoda]

Summary

Second phase of the initial catch-up per SYNC.md (Scoring area). Compared the cumulative pyscn origin/main (8650522) diff against jscan's current state for domain/analyze.go / domain/system_analysis.go / domain/complexity.go and ported the missing scoring changes.

Ported changes

pyscn commit	Change	jscan files changed
0886cfb, 760bc87, 68b6d2b	Switch the duplication metric to K-Core clone-group density (groups per 1000 lines × 20, capped at 10%) with a 0–10% penalty scale	domain/analyze.go, service/output_formatter.go
333c9ac, 9c84a3d	Soften CBO coupling calibration (medium weight 0.5→0.3, saturation at weighted ratio 0.40). Introduce the CouplingMediumWeight / CouplingSaturationRatio constants	domain/analyze.go
3c9927c	Use compliance directly as the ArchitectureScore (98% compliance = 98 points)	domain/analyze.go
e6b9920	Expose WeightedViolations to make ComplianceScore reproducible (ported for type alignment)	domain/system_analysis.go
2cc013c	Report module-scope code as <module> instead of __main__ (introduce the domain.ModuleFunctionName constant and route all internal sentinel comparisons through it)	domain/complexity.go, internal/analyzer/cfg_builder.go, service/complexity_service.go, service/dead_code_service.go, service/dead_code_aggregate.go

Mirroring pyscn's test updates (expected-value changes in analyze_test.go), added duplication / CBO / architecture score tests to domain/analyze_test.go, adapted to jscan's formulas.

Skipped changes (recorded in SYNC.md pending changes)

• LCOM cohesion scoring (54698b6, 57fcc66) — depends on the unported LCOM4 feature
• Cognitive complexity / RawMetrics (8aa6d21, ce469ce) — unported features
• Mock/synthetic fixture detection (15bd414) — pyscn-specific lint feature
• Suggestions (1eaf9f6) — depends on the unported domain/suggestion.go
• Analyze config plumbing (222d190, 779200a, 13644b4, e79a414, d6fd3da, ca75d12) — jscan's config-loader layer is structured differently, out of scope
• Architecture style presets / warn rules / neutral prefixes / cohesion-responsibility validation (ad79460, 1c9839d, 61f0431, 3708845, ddcf1c5) — jscan does not implement architecture validation (types are unused scaffolding)
• AbstractClassCount (f63540d) — Python ABC detection; jscan computes abstractness its own way in coupling_metrics.go from TS interfaces/abstract classes
• Django migration exclusion / Python default patterns (a33f1c7) — Python-specific

Also documented jscan's intentional divergences in the SYNC.md file-mapping notes (calculateComplexityPenalty uses the high/medium count ratio; calculateDeadCodePenalty uses a per-file rate) so future syncs don't overwrite them.

Verification

• make test passes
• Golden comparison on test-repos/ (latest-version, junk, express):
  • latest-version / junk: the only difference is architecture_score 100→0, exactly as the ported formula dictates (jscan does not implement architecture validation, so compliance=0; pyscn produces the same value when arch is disabled). This is a JSON-only field, not shown in HTML/CLI output
  • express: code_duplication_percentage 56.4→10.0 (the group-density metric's cap; the penalty is the full 20 both before and after, so the health score stays at 74/C), and architecture_score 100→0 (same as above). The ±a-few jitter in clone_pairs reproduces across re-runs of the same binary — pre-existing nondeterminism, unrelated to this change (clone detection is untouched). Function lists, complexity, and dead-code aggregates match exactly

🤖 Generated with Claude Code