Fix Mgxs::get_xs CHI_DELAYED indexing the wrong tensor when gout is null

[dallonby]

Description

Found while porting the OpenMC transport engine to Apple Silicon GPUs (Metal) — dallonby/openmc-metal — where flattening the MG data model for the device meant auditing these tensor layouts against the code that reads them.

The gout == nullptr branches of the CHI_DELAYED case in Mgxs::get_xs (src/mgxs.cpp) sum delayed_nu_fission where they should sum chi_delayed:

• delayed_nu_fission is rank 3 with shape [angle][delayed group][incoming group] (src/xsdata.cpp), but was indexed with four indices, (a, *dg, gin, g). tensor::Tensor::operator() performs no rank check, so the four-index read walks past the stride vector and reads out of bounds.
• In the dg == nullptr branch, the inner loop bound was delayed_nu_fission.shape(3), which returns 0 for a rank-3 tensor — so that branch silently returned 0 rather than a spectrum sum.

Both branches are meant to return chi_delayed summed over outgoing groups (mirroring the structure of the CHI_PROMPT case); they now read chi_delayed with its actual shape [angle][delayed group][incoming group][outgoing group] and loop over chi_delayed.shape(1) / shape(3).

Also fixed: the dimension comment for chi_delayed in include/openmc/xsdata.h documented the layout as [angle][in group][out group][delayed group], which does not match the allocation in src/xsdata.cpp ({n_ang, n_dg, n_g, n_g}). The comment now matches the code.

Reachability / impact: no current caller reaches the fixed branches — the only external CHI_DELAYED/CHI_PROMPT caller (src/random_ray/flat_source_domain.cpp) always passes a non-null gout, and the direct-index gout != nullptr path (which is correct) is untouched. So this changes no existing results; it removes silent out-of-bounds reads / silent zeros waiting for the first caller that asks for a group-summed delayed spectrum.

A possible follow-up hardening (not included here to keep this minimal): a debug-mode rank assertion in tensor::Tensor::operator(), which would have turned this into a loud failure. Happy to add a unit test exercising get_xs(CHI_DELAYED, gin, nullptr, ...) if that's preferred.

Checklist

• I have performed a self-review of my own code
• I have run clang-format on any C++ source files (if applicable)
• I have followed the style guidelines for Python source files (not applicable)
• I have made corresponding changes to the documentation (the xsdata.h layout comment)
• I have added tests that prove my fix is effective (not included — the fixed branches are currently unreachable from any caller; can add a C++ unit test on request)

🤖 Generated with Claude Code