Description
Regression tests produce inconsistent numerical results between local runs and GitHub Actions CI. The same tests pass locally but intermittently fail in CI due to small floating-point discrepancies in vibrational entropy calculations (including force/torque and combined FT matrices).
The failures are non-deterministic and vary slightly in magnitude across runs, suggesting environment-dependent numerical differences rather than a logical error in the implementation.
This behavior leads to regression mismatches in component-level and total entropy values, even when the underlying inputs and configuration are unchanged. The issue appears to be related to floating-point accumulation order, linear algebra backend differences, and/or multithreading in numerical libraries.
Description
Regression tests produce inconsistent numerical results between local runs and GitHub Actions CI. The same tests pass locally but intermittently fail in CI due to small floating-point discrepancies in vibrational entropy calculations (including force/torque and combined FT matrices).
The failures are non-deterministic and vary slightly in magnitude across runs, suggesting environment-dependent numerical differences rather than a logical error in the implementation.
This behavior leads to regression mismatches in component-level and total entropy values, even when the underlying inputs and configuration are unchanged. The issue appears to be related to floating-point accumulation order, linear algebra backend differences, and/or multithreading in numerical libraries.