test: compare JAX and Torch exports with the fitted equation

[MilesCranmerBot]

Why

test_feature_selection_custom_operators in test_jax.py and test_torch.py compared exported output with the target y. The fit stops at early_stop_condition=1e-5, which can accept an equation whose constants miss y by up to 0.0034 at 3 of 2,000 points. The check uses 3 decimals, so the tests failed on some runners with identical numbers each time, for example on master for #1390 and #1393 and on #1394.

Change

• JAX: compare the output of the checkpoint-loaded JAX export with the fitted equation, evaluated with sympy.lambdify on the selected features and a NumPy cos_approx.
• Torch: compare the Torch export with model.predict.

The tests now check export fidelity, which is their purpose, and no longer depend on how close the search gets to y.

Verification

• Both tests pass locally (Python 3.13, Julia 1.13).
• Changing the JAX mapping's x**6 / 720 to x**6 / 700 makes the JAX test fail (202 of 2,000 points mismatched), so it still catches a wrong export.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!