Add conditional platform BLAS linking for Rust backend

[igerber]

Summary

• Add conditional BLAS linking: Apple Accelerate on macOS, OpenBLAS on Linux via Cargo features
• Convert manual GEMV loop in sc_weight_fw_standard to ndarray::linalg::general_mat_vec_mul (dispatches to BLAS dgemv when enabled, falls back to ndarray's matrixmultiply kernel)
• Add rust_backend_info() diagnostic function reporting compile-time BLAS feature status
• Update CI workflows (publish.yml, rust-test.yml) with platform-specific --features flags and openblas-devel/libopenblas-dev install steps
• Add extern crate blas_src for linker flag propagation (required for linker-only crates)
• Windows builds unchanged: continue using pure Rust with no external dependencies

Methodology references (required if estimator / math changes)

• Method name(s): N/A — no methodology changes. This is purely a build/performance optimization.
• Paper / source link(s): N/A
• Any intentional deviations from the source (and why): None

Validation

• Tests added/updated:
  • tests/test_rust_backend.py: Added test_rust_backend_info test
  • rust/src/weights.rs: Added test_gemv_produces_correct_half_grad Rust unit test
• All 78 Rust backend equivalence tests pass
• All 60 SDID methodology tests pass
• All 37 SyntheticDiD estimator tests pass
• Local build verified with maturin develop --release --features accelerate on macOS ARM64
• rust_backend_info() correctly reports {'blas': True, 'accelerate': True, 'openblas': False} on macOS

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[igerber]

/ai-review

[github-actions]

🔁 AI review rerun (requested by @igerber)

Head SHA: 1dabaeab6a8a2999555ff11693717dd31978c14f

---

Overall assessment: ✅ Looks good

Executive Summary

• Affected method is SyntheticDiD/SC Frank–Wolfe weights (sc_weight_fw); the new Gram/standard loops preserve the registry-specified objective, gradient, and step formula, so methodology alignment looks intact. (docs/methodology/REGISTRY.md:L663-L711, rust/src/weights.rs:L220-L518)
• No changes to SE/variance, identification assumptions, or defaults; BLAS linking and rust_backend_info() are build/diagnostic only.
• Added tests cover Gram vs standard equivalence and GEMV correctness, reducing regression risk.

Methodology
Affected methods: SyntheticDiD unit/time weights (Frank–Wolfe sc.weight.fw). Cross-check against the Methodology Registry’s FW step and objective shows the updated loops implement the same half_grad, step size, and convergence rule; no deviations found. (docs/methodology/REGISTRY.md:L663-L711, rust/src/weights.rs:L220-L518)
Findings: None.

Code Quality

• P3: Impact: The comment “Zero heap allocations per iteration” is inaccurate because the periodic refresh reallocates ata_x every GRAM_REFRESH_INTERVAL, which can mislead performance expectations. Fix: either adjust the comment to note periodic allocations or refresh into a preallocated buffer (e.g., general_mat_vec_mul(1.0, &ata, &lam, 0.0, &mut ata_x)) to make the claim true. (rust/src/weights.rs:L220-L321)

Performance
No findings.

Maintainability
No findings.

Tech Debt
No findings.

Security
No findings.

Documentation/Tests
No findings.

Tests not run (not requested).