Investigate det_sign_exact regression from v0.4.2 to v0.4.3

[acgetchell]

Summary

det_sign_exact() appears to have regressed between v0.4.2 and v0.4.3. Investigate the behavioral or performance change, identify the first bad change, and fix or document the intended behavior before v0.4.4.

Current State

• det_sign_exact() is part of the exact feature path.
• Earlier exact determinant work and the det_sign_exact short-circuit are closed, but the observed regression suggests the current implementation may no longer match the v0.4.2 behavior.
• The regression needs confirmation against the released versions before choosing the fix.

Proposed Changes

• Reproduce the regression by comparing v0.4.2, v0.4.3, and current main.
• Identify whether the regression is correctness, performance, or both.
• Add a focused regression test that fails on the bad behavior.
• Fix det_sign_exact() while preserving exact-path invariants:
  • no silent precision loss
  • no non-finite propagation
  • agreement with det_exact() wherever both paths are defined
• Update documentation or benchmarks if the regression is performance-related.

Benefits

• Restores confidence in the exact determinant sign path.
• Protects the exact feature from release-to-release regressions.
• Keeps v0.4.4 focused on fixing known release-train regressions.

Implementation Notes

• Check changes around the Shewchuk-style f64 sign filter and Bareiss fallback.
• Compare benchmark and test behavior with the exact feature enabled.
• Useful validation commands:
  • cargo test --features exact --verbose
  • cargo bench --features exact or just bench-exact if performance is implicated