Weight Window Edge Case Fixes

[jtramm]

Fix two weight window application edge cases

Note: This PR requires commits from #3965, so definitely wait to review this PR until #3965 is merged first!

Summary

Two small fixes to how weight windows are applied during transport, both found while auditing apply_weight_window() during the MGXS bootstrapping work (#3965):

1. Zero lower bounds turn the weight window game off instead of demanding unbounded splitting.
2. The weight cutoff is normalized with the window bounds instead of being applied in absolute units inside an otherwise frame-normalized game.

Neither change affects any existing test results.

Change 1 — zero-bound cells play no game

A weight window lower bound of zero means no weight window information exists for that cell. OpenMC's weight window generators already encode this convention by marking cells with no tally information as -1 (which fails WeightWindow::is_valid()), and MCNP uses zero to disable the game per cell — from the MCNP 6.2 manual (WWN card, note 2): "If wij=0, the weight-window game is turned off in cell j for energy or time bin i and the weight cutoff game is turned on with a 1-for-2 roulette limit." Files imported with WeightWindowsList.from_wwinp pass those zeros through unchanged.

Previously a zero lower bound passed is_valid() and was treated as a real window: any particle entering the cell sat above the zero upper bound and split by max_split at every weight window checkpoint (weight/0 = inf), multiplying the population until terminated by absorption/leakage, the weight cutoff, or max_history_splits (default 10^7). Besides wwinp imports, zero-bound cells are theoretically (if extremely unlikely) reachable from FW-CADIS generation when a cell has a nonzero forward tally but a zero adjoint flux in some energy group (the inverse step maps those to exactly 0.0 rather than -1). A separate pathology affects particles born under a zero window: wgt_ww_born is set to the window midpoint of zero, producing an infinite normalization factor and (with the cutoff scaled, see Change 2) killing such particles at birth. Both misbehaviors are cured by playing no game.

The fix is to treat any non-positive lower bound as "no information": is_valid() now requires a strictly positive lower bound, making the zero case consistent with the generator's -1 convention and with MCNP's semantics.

Change 2 — weight cutoff normalized with the window

In PR #3459 - I added the ability for apply_weight_window() to rescale the window into the particle's birth-weight frame — lower, upper, and survival_weight are all multiplied by wgt_born / wgt_ww_born (and by ww_factor where applicable) — but weight_cutoff was compared in absolute units. The cutoff kill is the one biased operation in the weight window game (weight is deleted without roulette), justified by the discarded weight being negligible — and "negligible" is a statement relative to the window bounds, not an absolute scale. For particles born with weights far from the window normalization (e.g., strongly biased sources or unusual window normalizations), an absolute cutoff can sit inside the normalized windows and silently cull in-window particles. WeightWindow::scale() now scales the cutoff with the bounds, keeping it at the same relative depth below the window in every frame.

Testing

The fix is pinned by two invariant tests built on the existing weightwindows test problem, each expressing one consequence of "a non-positive lower bound means no information." Both are self-comparing (two runs of the same model with identical seeds must agree), so they need no reference files and no statistical tolerances, and each was verified to fail without Change 1. They cover the two distinct code paths a zero bound corrupts:

• test_zero_bound_windows_play_no_game: all-zero windows must transport identically to weight windows disabled. This covers particles born under a zero window, where the birth normalization (wgt_ww_born = window midpoint = 0) previously produced an infinite scale factor. Without the fix the tallies differ; with it the runs match. ~1.5 s.
• test_zero_and_negative_bounds_equivalent: windows with a zero-bound region must transport identically to the same windows with a -1 region, with particles born under valid windows encountering the region in flight. This covers the splitting pathology: without the fix this test does not complete within ten minutes (the zero region splits every particle by max_split at every checkpoint, even at the test model's max_history_splits of 200); with it the runs match. ~1.9 s.

The full weight window test family and the MGXS bootstrap regression test pass with no reference changes: the existing test window files use -1 (not 0) for no-information cells, and at the default cutoff of 1e-38 no particle weight approaches the cutoff in either frame.

Change 2 has no observable effect in any existing test by construction; it is a consistency fix whose behavior difference only appears for cutoffs or normalization frames far from the defaults. No test is added for it — constructing one would require a contrived frame, and the runtime consequence of the old behavior (a biased kill) is exactly the kind of silent statistical effect a fixed-seed test cannot meaningfully pin.

Checklist

• I have performed a self-review of my own code
• I have run clang-format (version 18) on any C++ source files (if applicable)
• I have followed the style guidelines for Python source files (if applicable)
• I have made corresponding changes to the documentation (if applicable)
• I have added tests that prove my fix is effective or that my feature works (if applicable)