Address TODO items: safe_inference migration, module splits, np.dot, and bug fixes

[igerber]

Summary

• Migrate ~32 inline inference sites to safe_inference() across 11 source files for consistent NaN-safe inference
• Split 3 oversized modules into sub-modules: trop.py → trop_results.py; imputation.py → imputation_results.py + imputation_bootstrap.py; two_stage.py → two_stage_results.py + two_stage_bootstrap.py
• Replace @ operator with np.dot() at ~30 observation-dimension matrix sites to avoid Apple M4 BLAS spurious RuntimeWarnings
• Fix diagnostics.py:785 SE returning 0.0 instead of NaN for undefined standard errors
• Fix test_two_stage.py::test_nan_propagation no-op test (was pass with no assertions)
• Add regression tests for staggered.py CI bug fix, np.dot warnings, and diagnostics SE behavior
• Update CLAUDE.md, TODO.md, and REGISTRY.md documentation

Methodology references (required if estimator / math changes)

• Method name(s): All estimators (inference computation consistency)
• Paper / source link(s): N/A — no methodology changes, only inference computation consolidation
• Any intentional deviations from the source (and why): TROP CI now uses t-distribution (was normal-distribution for CI but t-distribution for p_value — inconsistent). After safe_inference() migration, both use t-distribution. Documented in REGISTRY.md.

Validation

• Tests added/updated: test_diagnostics.py (SE=NaN regression), test_linalg.py (np.dot no-warnings), test_staggered.py (NaN-SE CI regression), test_two_stage.py (test_nan_propagation fix)
• Full test suite: 735+ tests passing across all affected modules (two_stage: 56, imputation: 76, trop: 78, staggered: 95, estimators: 142, diagnostics: 32, linalg: 66, sun_abraham: 47, utils: 83, triple_diff: 50)
• Backward compatibility verified: all re-exported imports work (from diff_diff.X import Y)

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall assessment: ✅ Looks good

Executive summary:

• Safe inference migration is consistent across the touched estimators, and NaN-gating now aligns with the library’s stated inference policy.
• TROP’s CI distribution change (normal → t) is implemented with df = max(1, n_treated_obs - 1) and is documented in the Methodology Registry.
• Module splits for imputation/two_stage/trop preserve re-exports and keep public API stable.
• Added tests cover NaN propagation and regression cases for CI/SE behavior and the np.dot warning fix.
• Only gap I see is a missing targeted test for the TROP CI distribution change.

Methodology

• No findings. TROP CI distribution change is documented and matches code in diff_diff/trop.py:1172 and diff_diff/trop.py:1746, consistent with docs/methodology/REGISTRY.md:945.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• P3 Impact: The TROP CI distribution change (t vs normal) is not explicitly regression-tested, so a future refactor could revert to inconsistent CI/p_value without detection. Concrete fix: add a small unit test in tests/test_trop.py that computes the CI for a small n_treated_obs and asserts t-based critical values are used. Location: diff_diff/trop.py:1172 and diff_diff/trop.py:1746, registry note docs/methodology/REGISTRY.md:945.

Tests not run (not requested).