Add axis-G Rust vs Python backend parity edge-case tests

[igerber]

Summary

Phase 2 silent-failures audit — axis-G (backend parity). Closes the coverage gap the audit flagged on three Rust-backed solver surfaces (findings #21, #22, #23). Test-only PR; any discovered divergences are marked xfail(strict=True) and logged to TODO.md as P1 follow-ups rather than fixed in-scope.

• Update roadmap with current implementation limitations #21 solve_ols skip-rank-check parity — three tests in TestSolveOLSSkipRankCheckParity covering mixed-scale, near-singular full-rank, and rank-deficient designs under skip_rank_check=True on HC1. All pass at rtol=1e-6, atol=1e-8 on fitted values; no Rust-side code change needed.
• Implement CallawaySantAnna covariate adjustment for conditional parallel trends #22 compute_synthetic_weights parity — three tests. Near-singular Y'Y passes. Extreme Y scale (1e9) and lambda_reg variations xfail(strict=True) — root cause: Rust path is Frank-Wolfe, Python fallback is projected-gradient descent. Same QP, different simplex vertices.
• Address code review feedback for CallawaySantAnna covariates #23 TROP Rust grid-search + bootstrap parity (@pytest.mark.slow) — grid-search ATT on rank-deficient Y xfail(strict=True) (~6% divergence). Bootstrap SE under fixed seed xfail(strict=True) (~28% divergence due to RNG-backend mismatch: Rust rand crate vs numpy default_rng).

The four xfail(strict=True) markers baseline the gaps so we get notified if/when the algorithms align. Three P1 follow-up entries added to TODO.md.

Methodology references

• Method name(s): N/A - no methodology changes
• Paper / source link(s): N/A
• Any intentional deviations from the source (and why): None

Validation

• Tests added/updated: 3 new test classes in tests/test_rust_backend.py (8 tests total: 4 pass, 4 xfail).
• Backtest / simulation / notebook evidence: N/A (test-only PR)
• Local verification:
  • pytest tests/test_rust_backend.py -m 'not slow' → 72 passed, 14 deselected, 2 xfailed
  • pytest tests/test_rust_backend.py::TestTROPRustEdgeCaseParity -m '' → all pass (2 xfailed as expected)
• DIFF_DIFF_BACKEND coverage: tests patch HAS_RUST_BACKEND=False to force Python path, matching the pattern already established by test_trop_global_solver_parity_no_lowrank (test_rust_backend.py:1687).

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall Assessment

✅ Looks good

No unmitigated P0/P1 findings in the changed files. This PR is limited to backend-parity tests plus TODO tracking; it does not change estimator math, weighting, variance/SE code, identification checks, or defaults.

Executive Summary

• The diff only touches tests/test_rust_backend.py:2115 and TODO.md:86; no production estimator implementation changed.
• Methodology cross-check is clean: TROP still follows the existing registry contract that variance is bootstrap-based and n_bootstrap >= 2, with inference flowing through the current implementation in docs/methodology/REGISTRY.md:1993, docs/methodology/REGISTRY.md:2017, and diff_diff/trop_global.py:842.
• The new tests target the intended backend gates: solve_ols(skip_rank_check=True) in diff_diff/linalg.py:621, synthetic-weight dispatch in diff_diff/utils.py:1180, and TROP global LOOCV/bootstrap branches in diff_diff/trop_global.py:688 and diff_diff/trop_global.py:966.
• The three newly added TODO entries are under the project’s “Tech Debt from Code Reviews” section at TODO.md:86, TODO.md:87, and TODO.md:88, so these parity gaps are now explicitly tracked rather than implicit.
• The slow TROP parity class is still exercised in automation: default pytest excludes slow tests via pyproject.toml:103, but Rust CI runs tests/test_rust_backend.py with -m '' in .github/workflows/rust-test.yml:145.
• I could not execute the tests in this review environment because pytest is not installed.

Methodology

• No findings. Affected surfaces are existing backend implementations only: solve_ols, compute_synthetic_weights, and TROP global parity. The PR does not alter estimator equations, assumptions, weighting, or SE formulas.

Code Quality

• No findings. The mock-based Rust/Python forcing pattern matches the existing style in tests/test_rust_backend.py:1571 and correctly patches the module-level backend gates used by the tested paths.

Performance

• No findings. The heavier TROP parity coverage is correctly isolated behind @pytest.mark.slow at tests/test_rust_backend.py:2317, so default suite runtime is unchanged.

Maintainability

• No findings. The new TODO entries are concrete, scoped to specific call sites, and tied directly to the strict xfail baselines.

Tech Debt

• Severity: P3
  Impact: The backend parity gaps are now explicitly tracked in TODO.md:86, TODO.md:87, and TODO.md:88, which is the right treatment for this test-only PR under the project’s deferred-work policy.
  Concrete fix: None required for approval; the follow-up work is already recorded.

Security

• No findings. The diff adds tests and TODO text only; no new secret-handling, I/O, or network-facing behavior appears.

Documentation/Tests

• No findings. xfail(strict=True) is the correct mechanism for these baselines because an eventual backend alignment will surface as XPASS instead of disappearing silently.
• Verification note: I was not able to run pytest locally in this environment because the executable is unavailable.