ORCA 6 ECP Molden: SIGSEGV in genP after spherical-basis back conversion

[Stardust0831]

Reproducer

ORCA 6.1 orca_2mkl -molden output for an ECP-containing closed-shell calculation triggers a deterministic SIGSEGV during loading.

• Method: wB97M-V def2-TZVP def2/J RIJCOSX gCP(DFT/TZ)
• ECP elements: I (def2-ECP, 25 effective charge); analogous failures occur for Pb and Cs/I systems.
• Minimal observed case: CGBA_HCl_FAI_sp_full.molden.input (38 atoms, 778 spherical basis functions, 778 MOs, 184 electrons, 92 occupied MOs).

Correct ECP metadata

The ORCA file contains:

[Pseudo]
I 38 25

A correct Multiwfn-style supplement is:

[Nval]
I 25

Multiwfn confirms Nuclear charge of I has been set to 25, so the ECP charge is parsed correctly. Replacing [Pseudo] with this correct two-column [Nval], or retaining both sections, does not avoid the crash.

Failure point

Both the existing 3.8-dev binary and the official 2026.7.15 Linux noGUI binary stop after:

Converting basis function information to GTF information...
Back converting basis function information from Cartesian to spherical type...
Generating density matrix...
forrtl: severe (174): SIGSEGV, segmentation fault occurred

The source path is fileIO.f90 ORCA Molden loading / spherical conversion followed by call genP; genP calls matmul_blas on CObasa without prior shape/content diagnostics.

Requested fix

Please add a minimal guard before the density-matrix BLAS call: validate the coefficient matrix dimensions and occupied-MO count, then produce a controlled diagnostic rather than SIGSEGV. A debug build with bounds checking should identify the exact invalid access if present.

This issue is being used to track a minimal noGUI Linux/Windows fix and CI artifact validation.

[Stardust0831]

Update: candidate fix has now been built and tested.

• Commit: fc570b61754ddc4fa9db78740c1166bb201b6478 (fix/orca-ecp-genp-safe-matmul).
• CI: https://github.com/Stardust0831/Multiwfn/actions/runs/29874076081 passed for Linux/macOS/Windows noGUI.
• The Linux noGUI release-candidate artifact was run on dxy4 against the original FAI ORCA 6 ECP Molden reproducer containing [Pseudo] I 38 25 plus correct [Nval] I 25. It returned 0 and wrote density.cub; the unpatched official and legacy binaries both SIGSEGV at Generating density matrix....
• A 60x60x60 density cube integrated to 183.7953 e for the 184-electron system (finite box/grid truncation), which also confirms the RHF density is not double-scaled.

The narrowly scoped change replaces only the RHF matmul_blas density-matrix construction with the mathematically equivalent intrinsic matmul after a column-count guard. This validates the workaround for the reproducer; it does not yet establish why the BLAS path faults after ORCA spherical-basis conversion.