# Analyzer Stabilization: const-correctness, cross-TU robustness, and app/CLI refactor

[SizzleUnrlsd]

Context

This change set improves analyzer stability on self-analysis runs, reduces inter-TU false negatives/false positives, and cleans up key architecture layers (app/CLI execution flow and C++ API quality).

Changes Delivered

1) Parameter debug binding refactor

• Introduced a dedicated ParameterDebugBinding component (parameter name/type/location + confidence).
• Centralized parameter debug binding logic instead of scattered local heuristics.
• Integrated binding usage in ConstParamAnalysis and diagnostic emission.

2) ConstParamAnalysis hardening

• Replaced boolean “may write” logic with a tri-state model:
  • NoWrite
  • Unknown
  • MayWrite
• Improved use of LLVM metadata (ReadOnly, ReadNone, WriteOnly, etc.).
• Added confidence scoring for findings.
• Suppressed non-actionable cases (artificial/forwarding/callable patterns for specific rvalue-ref scenarios).
• Made const-suggestion messages consistent with final resolved parameter binding.

3) Diagnostic emission improvements

• DiagnosticEmitter now uses parameter binding data (name + line/column).
• Confidence is propagated into diagnostics when available.

4) Cross-TU regression fix (resource summary cache)

• Fixed root cause: cross-TU cache contamination after filtered runs (--only-function, --only-file, etc.).
• Analysis filter context is now part of the resource cross-TU cache key.
• Result: no reuse of partial summaries across different analysis contexts.

5) App execution flow refactor

• Simplified runAnalyzerApp API (localized LLVMContext handling in execution path).
• Cleaned execution strategy boundaries (SharedModuleLoading vs direct loading) with simpler interfaces.
• Better separation of concerns between planning, execution, and output.

6) CLI SMT parsing cleanup

• Refactored SMT option handlers to a unified optional<error> result model.
• Reduced ad-hoc error paths and tightened parser consistency.

7) Const-correctness and C++20 hygiene

• Applied const-correctness updates across multiple analyses/pipeline paths.
• Updated CMake to include ParameterDebugBinding.cpp.
• Observed local config change: warnings-only=false.

Architecture Rationale

• Moving parameter debug resolution into a dedicated module improves reuse, testability, and maintainability.
• Fixing cache identity (key design) is a generic, durable solution compared to rule-specific workarounds.
• Separating plan/execution/output in the app layer reduces side effects and improves inter-TU pipeline evolvability.

Validation

• Build: OK (cmake --build build -j4)
• Repro (filtered run -> unfiltered run cache scenario): fixed
• Full test campaign: Passed 1658/1658 with --no-cache
• No test files were modified

Risks / Notes

• warnings-only=false changes default output behavior; confirm whether this is intended product behavior or local preference.
• Filter-aware cache partitioning increases cache entry count (expected tradeoff for correctness).

Acceptance Criteria

• No cross-TU diagnostic loss after filtered runs.
• Stable self-analysis behavior on regression suite.
• Green build and full tests without test changes.