fix: correct several neutron sampling / tally bugs

[steps-re]

Five independent correctness fixes found while auditing the neutron physics and tally code. Each is explained below with the reasoning.

1. Delayed neutrons are never sampled (transport/physics/neutron/native.py, fission())

xi = rng.lcg(particle_container_new)   # xi in [0, 1)
total = nu_p                            # prompt MULTIPLICITY (~2-3)
if xi > total:                         # never true
    prompt = False
    ...

xi is a uniform [0,1) sample but total is set to the prompt multiplicity nu_p (order 2-3), so xi > total is always false and the delayed-neutron branch is dead. The subsequent loop accumulates delayed fractions into total, confirming total is meant to be the prompt fraction. Fix: total = nu_p / nu.

2. Delayed emission time uses β instead of λ (same file)

decay_rate = mcdc_get.nuclide.neutron_fission_delayed_fractions(delayed_group, ...)
...
particle_new["t"] -= math.log(xi) / decay_rate

Sampling an exponential emission time needs the delayed-group decay constant λ, but this reads the delayed fraction β. A dedicated neutron_fission_delayed_decay_rates getter already exists; use it.

3. Swapped transverse-basis branches (transport/distribution.py, sample_direction())

The branch that builds 1/sqrt(wx^2+wy^2) ran when abs(wz) >= 0.999 (axis ~parallel to z, where wx^2+wy^2 -> 0, dividing by ~zero), while the fixed basis (1,0,0),(0,1,0) ran otherwise (where it is not orthonormal to a general axis). The bodies were swapped and the # Axis nearly parallel to z comment was on the wrong branch. Swapped so the fixed basis handles the pole and the general construction handles everything else.

4. Divide-by-zero for a -z reference (transport/distribution.py, sample_white_direction())

if nz != 1.0: lets nz == -1.0 (an ordinary -z normal) into the branch that computes B = sqrt(1 - nz**2) = 0, then C = Ac / B. Guard with abs(nz) != 1.0; the else branch (using ny) already handles both poles.

5. Tally polar_reference corrupted (object_/tally.py)

self.polar_reference /= polar_reference / np.linalg.norm(polar_reference)

self.polar_reference defaults to [0,0,1]; dividing it element-wise by the normalized input yields [0, nan, inf] for e.g. [1,0,0]. Every sibling filter assigns directly — assign the normalized vector.

Testing

These are localized logic fixes; I verified each against the surrounding code and the referenced getters/constants but did not run the full numba/MPI suite locally. Happy to adjust to your conventions.

[ilhamv]

Hi, @steps-re. Thanks for bringing these up and fixing them!

Some of the errors were introduced during the major refactoring, and the others were just not properly tested.

After a parametric testing, it is found that two regression tests already cover some of the errors (failing right now since they use wrong test answers):

• moving_source detects Error 3 (swapped transverse-basis branches in source sampling_direction)
• basic_weight_window detects Errors 1 and 2 (continuous-energy delayed neutron fission sampling; the use of time_boundary covers the incorrect delayed emission)
  (I went ahead and updated the failing regression test answers.)

Errors 4 and 5 are currently not covered by the tests. We can add a unit or regression test to cover those two. Do you want to add it as part of this PR, @step-re? Or, we can make a follow-up Issue for the need for the new test.

This is an important update. Once merged, we will make the patch release v0.14.2.

Thanks, @steps-re!

• Update the CHANGELOG.md

[ilhamv]

The need for tests to cover the fixed Errors 4 and 5 will be brought up again after #472 is resolved and we proceed with completing test coverage.