[Preview] [Testing] Single precision Minimizers

[scal444]

No description provided.

[scal444]

This is not ready for official review, I"ll be breaking it into chunks for PRs.

[greptile-apps]

Confidence Score: 4/5

The implementation appears safe to merge, with a non-blocking request to restore the reduced FULL-precision UFF regression coverage.

Findings

1.  Full coverage is reduced ▶

Summary

• Adds precision-aware CUDA storage, arithmetic, reductions, Hessian state, and conversion workspaces.
• Exposes PrecisionMode through the Python bindings and public optimization APIs.
• Extends force-field and minimizer tests across both precision modes.
• Documents precision behavior and the batched-backend requirement for SINGLE.

_{Reviews (1) · Last reviewed commit: "Remove precision implementation-history ..."}

[greptile-apps]

Full coverage is reduced

This now drops the first UFF validation molecule for both precision modes and reduces the fixture from five molecules to four, although the stated basin instability only explains excluding it from SINGLE. This weakens the existing default FULL regression coverage while the force-field and minimizer implementations are being generalized. Please retain this molecule for FULL and apply any precision-specific expectation only to SINGLE. The same unconditional removal also appears in tests/test_uff.cu:301-302.