Overall Assessment

✅ Looks good

Executive Summary

• No unmitigated P0/P1 issues found in the shared OLS scale-invariance fix.
• The methodology change is documented in docs/methodology/REGISTRY.md, so the deliberate deviation from raw unscaled numeric behavior is tracked rather than silent.
• The rank-0 IndexError pattern appears consistently fixed in the touched linalg indexing sites.
• One P3 documentation/tech-debt issue remains: the new scope caveat/TODO overstates unresolved TripleDifference exposure and understates some already-covered solve_ols consumers.
• I could not execute the test suite in this environment because project deps (numpy, pytest) are unavailable.

Methodology

• Severity: P3
  Impact: The new scope text says TripleDifference OR/DR nuisance solves still bypass solve_ols and remain non-equilibrated (docs/methodology/REGISTRY.md:L474-L474, docs/methodology/REGISTRY.md:L2101-L2101, TODO.md:L82-L82, CHANGELOG.md:L18-L18), but the OR fit path actually routes through solve_ols at diff_diff/triple_diff.py:L1438-L1448, so that path already inherits this PR’s scale-robust solver. The same scope summary also omits covered solve_ols consumers such as ImputationDiD and TwoStageDiD (diff_diff/imputation.py:L1220-L1227, diff_diff/two_stage.py:L2305-L2312).
  Concrete fix: Narrow the caveat/TODO to the genuinely local non-equilibrated solves, and make the covered-estimator list consistent across REGISTRY.md, TODO.md, and CHANGELOG.md.

• No P0/P1 methodology defects identified in the Python/Rust solve_ols change itself.

Code Quality

• No findings. The rank-0 empty-index fix was applied consistently in the touched indexing sites (diff_diff/linalg.py:L1099-L1109, diff_diff/linalg.py:L2734-L2739, diff_diff/linalg.py:L4133-L4176, diff_diff/staggered.py:L1332-L1335).

Performance

• No findings. The extra QR pass is only taken after a raw deficiency is detected, so the common full-rank path still pays for one QR.

Maintainability

• No additional findings beyond the scope-tracking issue above.

Tech Debt

• No blocker. The remaining non-equilibrated estimator-local paths are explicitly tracked in TODO.md; only the TripleDifference wording needs correction.

Security

• No findings.

Documentation/Tests

• No additional defects found in the added tests from static review.
• Validation note: I could not run the tests locally because this environment does not have the required Python packages installed.

🔁 AI review rerun (requested by @igerber)

Head SHA: 95b6d6cdeb0e1ee61e083c68fc496b870a91a9d3

Overall Assessment

✅ Looks good

Executive Summary

• No unmitigated P0/P1 issues found in the shared OLS scale-invariance fix or the rank-0 guardrails.
• The prior re-review scope-tracking issue appears resolved: REGISTRY.md, TODO.md, and CHANGELOG.md now correctly keep TripleDifference’s OR point-estimate path in scope while carving out only the estimator-local nuisance solves that still bypass solve_ols.
• The empty-index-array fix was applied consistently across the touched OLS/logit/Poisson/CallawaySantAnna sites.
• One P3 documentation issue remains: the changelog overstates what the mixed scale+collinearity regression test actually pins.
• I could not execute the tests locally because pytest and numpy are not installed in this environment.

Methodology

• No findings. This PR changes the shared OLS numerical backend, not a DiD estimand or variance formula, and the documented scope now matches the code: TripleDifference OR point estimates route through solve_ols, while CallawaySantAnna and StaggeredTripleDifference local nuisance solves remain explicitly deferred and tracked (docs/methodology/REGISTRY.md:L470-L474, docs/methodology/REGISTRY.md:L2099-L2101, TODO.md:L82-L83, diff_diff/triple_diff.py:L1438-L1448, diff_diff/staggered_triple_diff.py:L1556-L1563).

Code Quality

• No findings. The rank-0 empty-index fix was applied consistently across the touched linalg sites and the CallawaySantAnna helper (diff_diff/linalg.py:L1095-L1111, diff_diff/linalg.py:L2731-L2735, diff_diff/linalg.py:L4132-L4179, diff_diff/staggered.py:L1328-L1335).

Performance

• No findings. The second QR pass is only taken after raw QR already reports a deficiency, so the common full-rank path still exits after one QR before solving (diff_diff/linalg.py:L152-L170).

Maintainability

• No findings. REGISTRY.md, TODO.md, and the unreleased changelog are now aligned on the covered estimator surface and the remaining follow-up work (docs/methodology/REGISTRY.md:L470-L474, docs/methodology/REGISTRY.md:L2099-L2101, TODO.md:L82-L83, CHANGELOG.md:L18-L18).

Tech Debt

• No findings. The remaining non-equilibrated local nuisance fits and DR/OR IF rank-guard work are properly tracked in TODO.md, so they are mitigated deferrals rather than blockers (TODO.md:L82-L83).

Security

• No findings.

Documentation/Tests

• Severity: P3. Impact: CHANGELOG.md says the new regression coverage proves raw collinear drop-selection preservation “including the mixed scale+collinearity case,” but the mixed-case test does not assert that behavior; it only checks retention of the huge independent column, full rank of the kept subset, and valid kept-coefficient VCV (CHANGELOG.md:L18-L18, tests/test_linalg.py:L1491-L1549). Concrete fix: narrow the changelog text so it distinguishes the two guarantees: the well-scaled collinear case preserves raw drop selection, while the mixed-scale+collinearity case only pins retention of an identified full-rank subset with valid inference on kept coefficients.
• Validation note: I could not run the tests locally because this environment lacks pytest and numpy.

🔁 AI review rerun (requested by @igerber)

Head SHA: facfa1259fdd1cf68152063c761468825b44d777

Overall Assessment

✅ Looks good

Executive Summary

• No unmitigated P0/P1 issues found in the shared OLS scale-invariance change or the new rank-0 guardrails.
• The prior re-review scope issue is resolved: REGISTRY.md and TODO.md now consistently carve out the estimator-local CallawaySantAnna / StaggeredTripleDifference nuisance solves and the DR/OR IF rank-guard follow-up (docs/methodology/REGISTRY.md:L474-L474, TODO.md:L82-L83).
• The previous changelog overclaim about the mixed-scale test was partly fixed: the test description now distinguishes well-scaled collinearity from mixed scale+collinearity (CHANGELOG.md:L18-L18, tests/test_linalg.py:L1491-L1549).
• One P3 documentation issue remains: the registry and one changelog sentence still state the drop-selection guarantee too broadly for mixed scale+collinearity inputs (diff_diff/linalg.py:L172-L180, docs/methodology/REGISTRY.md:L79-L79, docs/methodology/REGISTRY.md:L471-L474, docs/methodology/REGISTRY.md:L1180-L1180, CHANGELOG.md:L18-L18).
• I could not run the tests locally because this environment does not have numpy, scipy, or pytest.

Methodology

• No findings. This PR changes the shared linear-algebra implementation, not a DiD estimand or variance formula, and the remaining estimator-local nuisance / IF gaps are explicitly documented and deferred rather than silently changed (diff_diff/linalg.py:L94-L180, diff_diff/linalg.py:L294-L312, docs/methodology/REGISTRY.md:L79-L79, docs/methodology/REGISTRY.md:L474-L474, TODO.md:L82-L83).

Code Quality

• No findings. The empty-index-array fix was applied consistently across the touched rank-deficiency sites in OLS, logit, Poisson, and the CallawaySantAnna helper (diff_diff/linalg.py:L1095-L1111, diff_diff/linalg.py:L2734-L2734, diff_diff/linalg.py:L4133-L4178, diff_diff/staggered.py:L1329-L1334).

Performance

• No findings. The extra equilibrated QR pass is only taken after the raw QR already reports a deficiency, so the common full-rank path still exits after the first rank check (diff_diff/linalg.py:L152-L170).

Maintainability

• No findings. The scope notes across code/docs/TODO are now materially aligned on which estimator surfaces inherit the shared solve_ols fix and which still use local nuisance solves (CHANGELOG.md:L18-L18, docs/methodology/REGISTRY.md:L474-L474, TODO.md:L82-L83).

Tech Debt

• No findings. The remaining uncovered local OR scale-robustness work and DR/OR IF rank-guard are properly tracked in TODO.md, so they are mitigated deferrals rather than blockers (TODO.md:L82-L83).

Security

• No findings.

Documentation/Tests

• Severity: P3. Impact: _detect_rank_deficiency() switches to the equilibrated pivot whenever rank_eq > rank_raw, and the mixed-scale regression test only asserts retention of an identified full-rank kept subset plus valid inference on kept coefficients; it does not pin preservation of the raw dropped column in mixed scale+collinearity cases. The current registry/changelog wording still says the re-check works “without changing the dropped-column selection” or “does not change which column is dropped in a genuinely collinear design,” which is too broad for that mixed case (diff_diff/linalg.py:L172-L180, tests/test_linalg.py:L1491-L1549, docs/methodology/REGISTRY.md:L79-L79, docs/methodology/REGISTRY.md:L471-L474, docs/methodology/REGISTRY.md:L1180-L1180, CHANGELOG.md:L18-L18). Concrete fix: narrow those sentences to the guarantee the code actually provides: raw drop selection is preserved for well-scaled / no-scale-disparity collinear designs, while mixed scale+collinearity guarantees only a scale-correct identified kept subset with valid inference on the kept coefficients.
• Validation note: I could not execute the new tests locally because numpy, scipy, and pytest are not installed in this environment.

🔁 AI review rerun (requested by @igerber)

Head SHA: 06755a59d7d833228d4dc7f7137370672236f7b9

Overall Assessment

✅ Looks good

Executive Summary

• No unmitigated P0/P1 issues found in the shared OLS scale-invariance repair or the rank-0 guardrails.
• Methodology scope is now largely aligned: the solve_ols-routed estimators are clearly covered, while the remaining CallawaySantAnna / StaggeredTripleDifference local nuisance-solve exposure is explicitly carved out and tracked in docs/methodology/REGISTRY.md:L470-L474, docs/methodology/REGISTRY.md:L2101-L2101, docs/methodology/REGISTRY.md:L2111-L2111, and TODO.md:L82-L83.
• The prior re-review documentation issue is only partially resolved: the main registry note and changelog now qualify the guarantee to well-scaled collinear designs, but the shorter MultiPeriodDiD and TwoWayFixedEffects summaries still overstate it in a way that contradicts the new mixed scale+collinearity behavior.
• The empty-index/rank-0 fix was applied consistently across the touched rank-deficiency sites in OLS, logit, Poisson, and CallawaySantAnna: diff_diff/linalg.py:L1095-L1111, diff_diff/linalg.py:L2731-L2735, diff_diff/linalg.py:L4132-L4176, diff_diff/staggered.py:L1328-L1335.
• I could not execute the new tests locally because pytest is not installed in this environment.

Methodology

• No findings. Affected methods are the shared solve_ols consumers (DifferenceInDifferences, MultiPeriodDiD, TwoWayFixedEffects, ImputationDiD, TwoStageDiD, TripleDifference). This is a numerical linear-algebra robustness fix, not a change to an estimand, identification condition, or variance formula, and the remaining out-of-scope nuisance solves are documented rather than silently changed: docs/methodology/REGISTRY.md:L79-L79, docs/methodology/REGISTRY.md:L470-L474, TODO.md:L82-L83.

Code Quality

• No findings. The rank-0 / empty-index pattern fix appears complete across the touched sites, and the new _equilibrated_lstsq() helper keeps the scale-robust solve centralized in one place: diff_diff/linalg.py:L294-L312.

Performance

• No findings. The extra equilibrated QR pass is only taken after the raw pass already reports a deficiency, so the common full-rank path still exits after a single QR check: diff_diff/linalg.py:L152-L170.

Maintainability

• No findings. The scope notes across code/docs/TODO are materially more aligned than in the previous review, especially around which estimators inherit the shared fix and which still use local nuisance solves.

Tech Debt

• No findings. The remaining non-solve_ols scale-robustness work and the DR/OR IF rank-guard are now explicitly tracked in TODO.md:L82-L83, so they are mitigated deferrals rather than blockers.

Security

• No findings.

Documentation/Tests

• Severity: P3. Impact: The short registry summaries for MultiPeriodDiD and TwoWayFixedEffects still compress the new behavior to “rank count only” and “without changing the dropped column,” but _detect_rank_deficiency() explicitly adopts the equilibrated pivot when equilibration repairs a scale-induced undercount. The new mixed scale+collinearity test also only guarantees an identified kept subset, not preservation of the raw dropped column. This is documentation drift, not a numerical defect: docs/methodology/REGISTRY.md:L206-L208, docs/methodology/REGISTRY.md:L279-L280, diff_diff/linalg.py:L172-L185, tests/test_linalg.py:L1491-L1549.
• Concrete fix: Rewrite those two summary notes to match the qualified wording already used in the main registry note and changelog: unchanged drop selection only for well-scaled genuinely collinear designs; mixed scale+collinearity may adopt a different scale-corrected identified subset: docs/methodology/REGISTRY.md:L79-L79, docs/methodology/REGISTRY.md:L470-L474, CHANGELOG.md:L18-L18.
• Validation note: I could not run the new tests locally because pytest is unavailable in this environment (pytest: command not found).

🔁 AI review rerun (requested by @igerber)

Head SHA: 5590771c27ea1341cf19c2105e960b51abb66916

Overall Assessment

✅ Looks good

Executive Summary

• No unmitigated P0/P1 findings in the shared OLS scale-invariance repair, Rust backend update, or rank-0 guardrails.
• The prior re-review documentation issue is resolved: the short MultiPeriodDiD and TwoWayFixedEffects registry notes now match the mixed scale+collinearity behavior documented in the main note at docs/methodology/REGISTRY.md:L208-L208 and docs/methodology/REGISTRY.md:L280-L280.
• Methodology impact is limited to shared solve_ols numerical linear-algebra behavior used by DifferenceInDifferences, TwoWayFixedEffects, MultiPeriodDiD, ImputationDiD, TwoStageDiD, and TripleDifference; no estimand, identification rule, or variance formula changed. Remaining estimator-local nuisance solves are explicitly scoped out and tracked at docs/methodology/REGISTRY.md:L470-L474 and TODO.md:L82-L83.
• The empty-index/rank-0 fix was applied consistently across the touched rank-deficiency sites in OLS, logit, Poisson, and the CallawaySantAnna helper at diff_diff/linalg.py:L1095-L1111, diff_diff/linalg.py:L2731-L2735, diff_diff/linalg.py:L4132-L4174, and diff_diff/staggered.py:L1328-L1335.
• Added regression coverage matches the change set: scale invariance, finite DiD ATT with a 1e8-scale covariate, mixed scale+collinearity, and rank-0 OLS/Poisson handling at tests/test_linalg.py:L1411-L1549 and tests/test_linalg.py:L2044-L2077.
• I could not execute the tests locally because pytest is not installed in this environment.

Methodology

• No findings. Affected methods are the shared solve_ols consumers, and the behavior change is a documented numerical-robustness repair rather than an estimand/SE/assumption change: diff_diff/linalg.py:L94-L190, diff_diff/linalg.py:L294-L312, rust/src/linalg.rs:L66-L166, docs/methodology/REGISTRY.md:L79-L79, docs/methodology/REGISTRY.md:L470-L474.

Code Quality

• No findings. _equilibrated_lstsq() centralizes the solve change, and the rank-0 integer-index guard was propagated consistently across the touched sites: diff_diff/linalg.py:L294-L312, diff_diff/linalg.py:L1095-L1111, diff_diff/linalg.py:L2731-L2735, diff_diff/linalg.py:L4132-L4174, diff_diff/staggered.py:L1328-L1335.

Performance

• No findings. The extra QR pass only occurs after raw deficiency is detected, so the common full-rank path still exits after a single QR check; the added solve-side work is column scaling around an already SVD-based solve: diff_diff/linalg.py:L152-L170, diff_diff/linalg.py:L1145-L1148.

Maintainability

• No findings. Code, changelog, registry, and TODO now agree on the mixed scale+collinearity behavior and the remaining out-of-scope nuisance solves: CHANGELOG.md:L18-L18, docs/methodology/REGISTRY.md:L208-L208, docs/methodology/REGISTRY.md:L280-L280, TODO.md:L82-L83.

Tech Debt

• No findings. The remaining CallawaySantAnna/StaggeredTripleDifference scale-robustness gap and the DR/OR influence-function rank-guard are explicitly tracked, so they are mitigated deferrals under the review rubric: TODO.md:L82-L83.

Security

• No findings.

Documentation/Tests

• No findings. The prior re-review documentation drift is addressed in the short registry summaries, and the added tests are aligned with the repaired behaviors: docs/methodology/REGISTRY.md:L208-L208, docs/methodology/REGISTRY.md:L280-L280, tests/test_linalg.py:L1411-L1549, tests/test_linalg.py:L2044-L2077.
• Validation note: not run locally (pytest: command not found).