Fix NaN EI/PI gradients for tiny positive std

[alex-rosen]

Summary

This PR fixes a numerical stability issue in gaussian_pi and gaussian_ei when predicted std is extremely small but still positive.

In that regime, the gradient path can hit inf * 0 (overflow in improve_grad, underflow in norm.pdf(scaled)), which may produce NaN gradients.

What changed

• In skopt/acquisition.py:

  • Keep acquisition value computation unchanged for std > 0.
  • In gradient mode, if |scaled| > 35, return a zero gradient (stable tail-limit behavior) instead of proceeding through an unstable numeric path.

• In skopt/tests/test_acquisition.py:

  • Add a regression test:
    • test_ei_pi_gradients_are_finite_for_tiny_positive_std
    • Uses a tiny-std surrogate (std = 1e-160) with non-zero std_grad
    • Verifies EI/PI values and gradients are finite.

Why this is correct

For extreme tails, the EI/PI gradient should tend to zero. Returning zero gradient in this regime preserves optimization behavior while avoiding NaNs from numerical overflow/underflow.

Validation

Ran:
bash
pytest -q skopt/tests/test_acquisition.py -k "tiny_positive_std"
Result:

• 2 passed

Result:- 2 passed