Remove LOOCV control observation subsampling from TROP

[igerber]

Summary

• Remove max_loocv_samples parameter from TROP estimator — LOOCV now sums over ALL control observations per Equation 5 of Athey et al. (2025)
• Parallelize inner LOOCV loops in Rust backend via Rayon fold/reduce to compensate for increased observation count
• Delete test_joint_python_loocv_subsampling test and update all call sites

Methodology references (required if estimator / math changes)

• Method name(s): TROP (Triply Robust Panel Estimator)
• Paper / source link(s): Athey, S., Imbens, G. W., Qu, Z., & Viviano, D. (2025). Triply Robust Panel Estimators. https://arxiv.org/abs/2508.21536
• Any intentional deviations from the source (and why): This PR removes a deviation. The previous max_loocv_samples subsampling was incorrectly attributed to the paper's footnote, but Footnote 2 only describes the two-stage coordinate descent search strategy. Equation 5 and Algorithms 1/2 explicitly sum over ALL control observations where W_it = 0.

Validation

• Tests added/updated: tests/test_trop.py, tests/test_rust_backend.py
• All 79 TROP tests pass (Rust backend: 80s, Python-only: ~49min)
• All 62 Rust backend equivalence tests pass
• LOOCV is now fully deterministic (no random subsampling)

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall assessment: ✅ Looks good
Executive Summary

• TROP LOOCV now evaluates all control observations in both Python and Rust paths; aligns with the Methodology Registry’s Equation 5 requirements.
• Rust LOOCV per‑observation loops are parallelized; note a minor performance/registry alignment concern around lack of short‑circuiting on failures.
• Tests updated for new Rust LOOCV signatures and deterministic behavior; subsampling‑specific coverage removed.

Methodology

• No findings. Affected method: TROP LOOCV tuning; registry Eq. 5 (sum over all D==0 controls; any fit failure ⇒ Q(λ)=∞) matches the implementation.

Code Quality

• No findings.

Performance

• P3 | Impact: For λ combinations with any failed fit, the new parallel loops still evaluate all controls before returning Q(λ)=∞, adding avoidable work on failure‑prone grids and diverging from the registry’s “early return” note. | Fix: Add early‑exit via try_fold/try_reduce + AtomicBool (or similar) to short‑circuit once a failure is detected, or update the registry wording if full evaluation is intended. (rust/src/trop.rs:L448-L495, rust/src/trop.rs:L1398-L1441)

Maintainability

• No findings.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• No findings.

Open Questions

• Assumption: Paper text not in repo; methodology check relies on docs/methodology/REGISTRY.md and in‑code docstrings.

Change Summary

• Removed max_loocv_samples and subsampling; LOOCV uses all controls in Python and Rust paths.
• Rust LOOCV per‑observation scoring is parallelized.
• Tests updated for new Rust LOOCV signatures and deterministic behavior.