Overall Assessment
✅ Looks good

Executive Summary

• This is a tracker/documentation-only PR; I found no methodology or implementation regressions.
• The methodology-facing removals and rewrites are consistent with the current registry and code: shipped weighted CR2-BM support, completed vcov_type threading, retained {"hc1"} narrowing for IF/GMM estimators, and the narrowed HAD / local-linear follow-up rows all check out.
• No P0/P1 issues found.
• One minor P3 documentation drift remains: TODO.md still cites REGISTRY:2382, but the relevant HAD sup-t singleton-strata deviation now lives at docs/methodology/REGISTRY.md.
• I could not run pytest in this environment because pytest is not installed; validation here was static against source and existing test files.

Methodology

• No findings. The rewritten TODO.md “Standard Error Consistency” summary matches the registry’s variance taxonomy in docs/methodology/REGISTRY.md, the shipped weighted CR2-BM note in docs/methodology/REGISTRY.md, and existing threading coverage for TwoStageDiD in tests/test_two_stage.py.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• Severity: P3
  Impact: TODO.md says the Stute singleton-strata gap is “the same gap as the HAD sup-t deviation at REGISTRY:2382,” but docs/methodology/REGISTRY.md is now in the TROP section. The relevant HAD note is docs/methodology/REGISTRY.md. This does not affect behavior, but it weakens the accuracy of an audit/pruning PR.
  Concrete fix: Replace the brittle absolute line reference with the current HAD location, or better, use a section title/anchor instead of a raw line number.

Tech Debt

• No findings. The deleted helper/weights rows are consistent with current helper/wrapper surfaces, including _resolve_survey_for_fit in diff_diff/survey.py and the documented weighted local-linear gap in diff_diff/local_linear.py.

Security

• No findings.

Documentation/Tests

• No approval-affecting findings. The narrowed HAD pretest follow-up row matches the current registry/tests (docs/methodology/REGISTRY.md, tests/test_had_pretests.py), and the narrowed local-linear row matches current docs/tests (diff_diff/local_linear.py, tests/test_bias_corrected_lprobust.py).

🔁 AI review rerun (requested by @igerber)

Head SHA: 770af785917f66d2da6791f5cd917239ea2aaeb2

Overall Assessment
✅ Looks good

Executive Summary

• This is still a tracker-only PR. I found no unmitigated P0/P1 issues in the changed lines after cross-checking the live code and Methodology Registry.
• The prior review’s stale TODO-to-registry pointer is fixed: the HAD pretests row now cites the current Stute calibration section by name at TODO.md:128, consistent with docs/methodology/REGISTRY.md:2733.
• The deleted TwoStageDiD threading row is stale as claimed: vcov_type is already threaded, exposed in params/results, and regression-tested in diff_diff/two_stage.py:1311, diff_diff/two_stage.py:2067, diff_diff/two_stage.py:3262, and tests/test_two_stage.py:2020.
• The other pruned tracker rows also match shipped state: weighted WLS-CR2 is documented as lifted in docs/methodology/REGISTRY.md:2827, weighted bias_corrected_local_linear is live in diff_diff/local_linear.py:1019 and tested in tests/test_bias_corrected_lprobust.py:563, and the upstream survey-metadata dedup now lives in diff_diff/survey.py:1232 with only the post-cluster-injection recompute remaining in diff_diff/spillover.py:3018.
• One minor P3 wording nit remains in the rewritten Standard Error Consistency paragraph: it attributes TwoStageDiD’s permanent {"hc1"} narrowing to “influence-function variance,” while the registry documents a distinct Gardner GMM-meat rationale at TODO.md:231, docs/methodology/REGISTRY.md:348, and docs/methodology/REGISTRY.md:1408.

Methodology

• Severity: P3
  Impact: TODO.md:231 correctly says TwoStageDiD is permanently narrow to {"hc1"}, but the rationale is slightly overstated: the four IF-based estimators are narrow because of IF-based variance, while TwoStageDiD is narrow because its Gardner GMM-corrected meat has no single cross-stage hat-matrix analogue for classical/hc2/hc2_bm. The conclusion is right; the audit trail is just less precise than the registry. See docs/methodology/REGISTRY.md:348, docs/methodology/REGISTRY.md:1408, and diff_diff/two_stage.py:3262.
  Concrete fix: Split the rationale in that sentence, e.g. “the four IF-based estimators … are permanently narrow to {"hc1"} per their IF variance, and TwoStageDiD is likewise permanently narrow because its GMM-corrected meat has no derived HC2/CR2/classical analogue.”

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings. The previous TODO→REGISTRY drift called out in the last review is resolved in the changed file at TODO.md:128.

Tech Debt

• No findings. The removed rows now reflect shipped state rather than hiding open work, per docs/methodology/REGISTRY.md:2827, diff_diff/local_linear.py:1019, and diff_diff/survey.py:1232.

Security

• No findings.

Documentation/Tests

• No findings. Static review only; I did not run the test suite in this environment.

🔁 AI review rerun (requested by @igerber)

Head SHA: 68c0026e1f2e92ffdba867be97037b1673c2c4e3

Overall Assessment
✅ Looks good

Executive Summary

• This is still a tracker-only PR. I found no unmitigated P0/P1 issues in the changed lines.
• The previous review’s only methodology nit is addressed: the revised Standard Error Consistency paragraph now separates the four IF-based estimators from TwoStageDiD’s distinct Gardner-GMM rationale, consistent with TODO.md:L229-L231, docs/methodology/REGISTRY.md:L348-L357, and diff_diff/two_stage.py:L3262-L3295.
• The deleted standalone-vcov_type tracker rows are stale rather than hiding open work; the remaining per-estimator Conley follow-ups are still tracked at TODO.md:L103-L111.
• The removed bias_corrected_local_linear weights= row is also stale: weighted fits are live, and the remaining gap is correctly narrowed to weight-aware auto-bandwidth DPI at TODO.md:L223, diff_diff/local_linear.py:L1019-L1028, and tests/test_bias_corrected_lprobust.py:L563-L678.
• The narrowed HAD pretests row matches the registry and current guards: stratified survey support is shipped, while replicate weights and lonely_psu='adjust' + singleton strata remain the only open items at TODO.md:L128, docs/methodology/REGISTRY.md:L2633-L2633, docs/methodology/REGISTRY.md:L2729-L2745, and diff_diff/had_pretests.py:L3308-L3359.
• The deleted survey-metadata dedup row appears correctly closed: _resolve_survey_for_fit now owns the shared pre-fit path, and the remaining spillover.py recompute is a distinct post-PSU-injection update rather than the old duplication at diff_diff/survey.py:L1232-L1258, diff_diff/spillover.py:L2371-L2380, and diff_diff/spillover.py:L3018-L3034.

Methodology

• No findings. The changed prose in TODO.md:L229-L231 now matches the Methodology Registry and runtime guard for TwoStageDiD; the prior P3 wording issue is resolved.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings. The tracker deletions I spot-checked correspond to shipped state or genuine de-duplication, not disappearing active work.

Tech Debt

• No findings. Remaining deferred items are still tracked where expected, including standalone vcov_type="conley" follow-ups and the HAD pretest gaps at TODO.md:L103-L128.

Security

• No findings.

Documentation/Tests

• No findings. Static review only; I did not run the test suite in this environment.