Add Windows wheel builds using faer pure Rust linear algebra

[igerber]

Summary

• Replace ndarray-linalg (requires OpenBLAS/MKL) with faer (pure Rust linear algebra)
• Update Rust dependencies: pyo3 0.22, numpy 0.22, ndarray 0.16, faer 0.24, faer-ext 0.7
• Migrate linalg.rs SVD/LU to faer equivalents
• Migrate trop.rs soft_threshold_svd() to faer SVD
• Update all Rust files for PyO3 0.22 API (Bound<> return types)
• Add Windows to rust-test.yml CI matrix
• Add Windows wheel build job to publish.yml

Methodology references (required if estimator / math changes)

• Method name(s): OLS solver (SVD-based), matrix inversion (LU), nuclear norm soft-thresholding (SVD)
• Paper / source link(s): N/A - implementation change only, no methodology change
• Any intentional deviations from the source (and why): None - faer provides equivalent SVD/LU operations

Validation

• Tests added/updated: No new tests needed - all 62 existing Rust backend equivalence tests pass
• Full test suite: 1091/1091 tests passed
• Build verification: maturin build --release succeeds

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall Assessment

• ⚠️ Needs changes

Executive Summary

• P1: The new faer SVD path in OLS breaks underdetermined designs (n<k), so regression-based estimators can panic or return invalid β.
• TROP soft-threshold SVD changes still align with the Methodology Registry; no deviations found there.
• Tests don’t cover the new n<k SVD path, so this regression could slip into releases.

Methodology

• Severity: P1 | Impact: Under-determined designs (n<k) now hit a shape mismatch in the SVD pseudoinverse path, so OLS can panic or return invalid coefficients; this violates the registry’s rank-deficient SVD handling for regression-based estimators (DifferenceInDifferences, MultiPeriodDiD, TwoWayFixedEffects, SunAbraham). | Concrete fix: Use r=min(n,k); build s_inv_uty with length r and compute β via V (k×r) × s_inv_uty (or request full Vᵀ from faer), plus an explicit guard for n<k. | Location: rust/src/linalg.rs:121, rust/src/linalg.rs:132

Code Quality

• No issues found.

Performance

• No issues found.

Maintainability

• No issues found.

Tech Debt

• No issues found.

Security

• No issues found.

Documentation/Tests

• Severity: P2 | Impact: The faer-based OLS path lacks a regression test for underdetermined matrices, so the n<k failure mode can recur unnoticed. | Concrete fix: Add a unit test in rust/src/linalg.rs that exercises n<k (and, if possible, a Python rust-backend equivalence test for that case). | Location: rust/src/linalg.rs:336

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 5ddc8b17013d839df966dcbd557673655e0af0b2

---

Overall assessment: ⚠️ Needs changes

Methodology

• P1 | Impact: LU-based inversion only checks for NaN/Inf, so singular or near‑singular X'X can yield finite but invalid vcov/SEs; this conflicts with the Registry’s rank‑deficient handling for OLS‑backed estimators (DiD/TWFE/etc.). | Fix: add an explicit invertibility/rank check (pivot threshold or is_invertible/rank if available) and raise on singular X'X, or restore Cholesky+LU with failure detection. | Ref: rust/src/linalg.rs:L286

Code Quality

• No findings.

Performance

• P3 | Impact: Per‑element ndarray↔faer conversions and manual reconstruction loops add overhead in OLS SVD and TROP soft‑thresholding; may regress large‑matrix performance vs prior BLAS path. | Fix: use faer-ext ndarray interop or keep computations in faer and convert outputs once. | Ref: rust/src/linalg.rs:L63, rust/src/trop.rs:L770

Maintainability

• No findings.

Tech Debt

• P3 | Impact: faer-ext is declared but unused in Rust sources, adding dependency and build surface. | Fix: remove faer-ext from rust/Cargo.toml or adopt it for ndarray conversion. | Ref: rust/Cargo.toml:L18

Security

• No findings.

Documentation/Tests

• P2 | Impact: No regression test covering singular X'X handling after the LU switch; silent failures could slip through. | Fix: add a unit test that passes a singular design into invert_symmetric or compute_robust_vcov and asserts a ValueError. | Ref: rust/src/linalg.rs:L286

Open Questions/Assumptions

• None.

Executive Summary

• P1 risk: singular X'X may now produce finite but invalid SEs unless LU invertibility is explicitly checked.
• Add a regression test for singular X'X to lock in rank‑deficient behavior.
• Minor perf regression risk from per‑element conversions in OLS/TROP paths; consider faer-ext interop.
• Minor tech debt: faer-ext currently unused; remove or leverage it.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 2341e4f529baaf8cd49bb2567b60a005738c4e1c

---

Overall Assessment

• ⚠️ Needs changes

Methodology

• P0 Impact: bootstrap_trop_variance and bootstrap_trop_variance_joint return se=0.0 when fewer than 2 bootstrap estimates, violating NaN propagation and risking misleading inference in TROP; Concrete fix: return f64::NAN (and ensure downstream inference fields are NaN or force Python fallback) when bootstrap_estimates.len() < 2; Locations: rust/src/trop.rs:L1038-L1054, rust/src/trop.rs:L1703-L1719.
• P1 Impact: New LU residual gating throws on near‑singular but full‑rank X'X, changing default behavior from the registry’s rank‑deficient warn/NA path and causing Rust solve_ols to hard‑fail without a Python fallback; Concrete fix: align instability threshold with the rcond=1e-07 rank check or return NaN vcov instead of error (and/or catch Rust errors in _solve_ols_rust to fall back to NumPy/QR); Locations: rust/src/linalg.rs:L286-L351, rust/src/linalg.rs:L140-L205, diff_diff/linalg.py:L247-L320.

Code Quality

• No issues found.

Performance

• P2 Impact: Residual verification multiplies A * A^{-1} on every vcov computation, adding an extra O(k^3) matmul per solve that can be costly with large dummy‑expanded designs; Concrete fix: gate the check behind a debug flag, use a cheaper condition‑number estimate from LU, or only run the residual check when pivots indicate instability; Location: rust/src/linalg.rs:L326-L351.

Maintainability

• No issues found.

Tech Debt

• No issues found.

Security

• No issues found.

Documentation/Tests

• No issues found.

Executive Summary

• P0: TROP bootstrap SE returns 0.0 when insufficient samples; should be NaN or trigger fallback to avoid misleading inference.
• P1: LU residual gating can hard‑fail full‑rank regressions and bypass the registry’s warn/NA behavior; align thresholds or add fallback.
• P2: Added residual check incurs an extra O(k^3) matmul per vcov; consider gating for large models.
• Otherwise, the faer SVD/soft‑threshold logic appears aligned with registry formulas.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: f3b1566ad53f105b4efb7525db6fafce3fa757ff

---

Overall Assessment: ⚠️ Needs changes

Executive Summary

• P1: solve_ols(skip_rank_check=True) can return None when the Rust backend flags numerical instability, breaking the documented return contract (diff_diff/linalg.py:482).
• P2: New "numerically unstable" inversion errors from Rust propagate through compute_robust_vcov without fallback, so estimators that call it directly can now hard‑fail (diff_diff/linalg.py:752, rust/src/linalg.rs:367, diff_diff/estimators.py:436).
• P3: Registry/docstring cross‑check for OLS HC1/cluster SEs and TROP soft‑threshold SVD shows no methodology deviations (docs/methodology/REGISTRY.md:53, docs/methodology/REGISTRY.md:482, rust/src/linalg.rs:1, rust/src/trop.rs:1).

Methodology

• Severity: P3 | Impact: OLS‑based estimators (DiD/MultiPeriodDiD/TWFE) and TROP remain aligned with registry requirements (HC1/cluster SEs; soft‑threshold SVD); no deviations detected. | Concrete fix: None. (docs/methodology/REGISTRY.md:53, docs/methodology/REGISTRY.md:482, rust/src/linalg.rs:1, rust/src/trop.rs:1)

Code Quality

• Severity: P1 | Impact: solve_ols(skip_rank_check=True) may return None on Rust instability, which can crash tuple unpacking or leak None into estimator pipelines. | Concrete fix: In the skip_rank_check branch, detect None and fall back to _solve_ols_numpy to preserve the function’s return contract. (diff_diff/linalg.py:482)
• Severity: P2 | Impact: compute_robust_vcov now surfaces "numerically unstable" Rust errors without fallback, potentially breaking inference paths that use it directly (e.g., wild bootstrap storage). | Concrete fix: In compute_robust_vcov, treat "numerically unstable" like "Matrix inversion failed" and fall back to _compute_robust_vcov_numpy before raising. (diff_diff/linalg.py:752, rust/src/linalg.rs:367, diff_diff/estimators.py:436)

Performance

• Severity: P3 | Impact: Manual element‑wise copies and triple loops in faer SVD conversion/reconstruction may be slower for large matrices. | Concrete fix: Consider using faer’s matmul/conversion helpers to avoid nested loops where possible. (rust/src/linalg.rs:85, rust/src/trop.rs:837)

Maintainability

• Severity: P3 | Impact: No maintainability issues found in this diff. | Concrete fix: None.

Tech Debt

• Severity: P3 | Impact: No tech‑debt concerns identified in this diff. | Concrete fix: None.

Security

• Severity: P3 | Impact: No security issues found. | Concrete fix: None.

Documentation/Tests

• Severity: P3 | Impact: New Rust instability fallback paths are not covered by tests; regressions could slip in as faer behavior evolves. | Concrete fix: Add Rust‑backend tests that trigger instability and assert fallback behavior (and that skip_rank_check=True still returns valid tuples). (tests/test_linalg.py:303, diff_diff/linalg.py:482, diff_diff/linalg.py:752)

Open Questions/Assumptions

• Do you want compute_robust_vcov to always fall back to NumPy on numerical instability, or should it be a hard error for ill‑conditioned X'X?

Change Summary

• Replaced BLAS‑backed ndarray‑linalg with faer for Rust SVD/LU and added numerical‑instability checks/fallback hooks.
• Expanded CI to build/test Windows wheels and updated workflows accordingly.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 1a00fb9ee17fab5863fd02c1ff8a87694483f7bd

---

Overall assessment: ⚠️ Needs changes

Executive Summary

• Methodology cross-check: OLS solver (SVD + robust vcov) and TROP soft-threshold SVD remain aligned with docs/methodology/REGISTRY.md and Athey et al. (2025).
• P2: compute_robust_vcov silently falls back on Rust numerical-instability errors, which can mask unstable SEs.
• P3: No explicit test coverage for the new numerical-instability path.

Methodology

• Severity: P3 | Impact: None (alignment confirmed) | Fix: None — OLS SVD/robust SE formulas and TROP soft-threshold SVD remain consistent with docs/methodology/REGISTRY.md, rust/src/linalg.rs:44, and rust/src/trop.rs:784.
• Severity: P2 | Impact: compute_robust_vcov returns fallback SEs without warning when Rust reports numerical instability, so inference can appear valid despite ill-conditioned X'X | Fix: emit a warning or surface the error on fallback in diff_diff/linalg.py:764.

Code Quality

• No findings.

Performance

• No findings.

Maintainability

• No findings.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• Severity: P3 | Impact: The new LU instability path is untested, so warning/fallback behavior could regress | Fix: add a regression test that triggers Matrix inversion numerically unstable in rust/src/linalg.rs:365 and asserts the warning/fallback behavior in diff_diff/linalg.py:764.

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 69fde20457939a36d70c59245a149e93421f7147

---

Overall assessment: ✅ Looks good

Methodology

• P3: No deviations found versus docs/methodology/REGISTRY.md for SVD-based OLS, HC1/cluster vcov, or TROP soft-threshold SVD. Impact: N/A. Fix: N/A.

Code Quality

• P3: _solve_ols_rust only falls back on "numerically unstable"/"singular" errors; faer SVD failures raise a ValueError and bypass fallback. Impact: rare ill-conditioned inputs could hard-fail instead of using the Python backend. Fix: treat "SVD computation failed" as a fallback trigger (warn + return None). (diff_diff/linalg.py:307, rust/src/linalg.rs:70)

Performance

• P3: No performance regressions identified in this diff. Impact: N/A. Fix: N/A.

Maintainability

• P3: No maintainability issues identified in this diff. Impact: N/A. Fix: N/A.

Tech Debt

• P3: No tech-debt issues identified in this diff. Impact: N/A. Fix: N/A.

Security

• P3: No security issues identified in this diff. Impact: N/A. Fix: N/A.

Documentation/Tests

• P3: Backend swap to faer adds only a warning/fallback test; no explicit parity/regression tests for OLS coefficients/vcov or TROP soft-threshold SVD. Impact: numerical drift across platforms could slip through. Fix: add small deterministic parity tests comparing Rust vs NumPy/SciPy outputs. (tests/test_linalg.py:523)

Open Questions

• Do you want a small cross-platform parity test suite specifically for faer SVD/LU vs SciPy to guard against numeric drift?

Executive Summary

• No methodology deviations detected vs the registry for OLS/robust vcov or TROP soft-thresholding.
• Minor reliability gap: faer SVD failures bypass the Python fallback; consider mapping that error into the fallback path.
• Test gap: add parity tests for faer vs NumPy/SciPy to catch numerical drift across platforms.