QL: intermediate propagator join discards DGBSV failure details

[krystophny]

Reproduction

Build NEO-2-QL at 4b00b887b6a9c37928b6575130b9ad494d2cdc2b
with GNU bounds checks:

cmake -S . -B /tmp/neo2-debug -G Ninja \
  -DCMAKE_BUILD_TYPE=Debug \
  -DCMAKE_Fortran_FLAGS='-fcheck=all -fbacktrace -g -O0' \
  -DLIBNEO_REF=13387c26305fec88158ff109516a10729e2df206
cmake --build /tmp/neo2-debug --target neo_2_ql.x

Run one electron species on a HELCORE-type l=1 Boozer geometry with the
fileless, driven non-axisymmetric path:

&ntv_input
  isw_ntv_mode = 0
  isw_qflux_NA = 1
  in_file_pert = 'none'
  isw_m_phi_input = 1
  m_phi_input = -1
  isw_ripple_solver = 3
  isw_hel_drive = 1
  m_theta_hel = 1
  hel_phim_re = 1.0d0
/

The case uses iota=0.99, nperiod=1000, phi_split_mode=1,
phi_place_mode=1, bin_split_mode=0, max_solver_try=10, and one MPI rank.
It triggers normal ierr=3 mode-3 refinement retries, completes the local
propagator solves, and fails during intermediate propagator joining.

Expected

After successful local solves, the global propagators join and only final,
globally joined transport output is accepted. If the joining matrix is
singular or an argument is invalid, the diagnostic retains the LAPACK status,
join tags, and dimensions and exits through the normal MPI finalization path.

Actual

The run stops after the local solves:

solving completed
Error from Joining, ierr=           2
Note: The following floating-point exceptions are signalling: IEEE_INVALID_FLAG

The control path is:

• NEO-2-QL/propagator.f90:1403: intermediate join_ripples_interface
• NEO-2-QL/propagator.f90:1830: join_ripples_nn
• NEO-2-QL/join_ripples_int.f90:307-308: join_ripples
• NEO-2-QL/join_ripples.f90:138: gbsv
• NEO-2-QL/join_ripples.f90:140-142: every nonzero LAPACK info is reduced
  to ierr=2

COMMON/lapack_band.f90 calls DGBSV, where info<0 means an illegal
argument and info>0 means an exact zero U pivot. The actual value is lost,
so the present diagnostic cannot distinguish these cases. The IEEE flag is
reported at the explicit STOP; it is not established as the cause.

The run leaves qflux_ntv.dat and qflux_symm.dat, but those files are
written inside local ripple solves before global joining. They are partial
outputs and are not trustworthy as final transport results.

An optimized build diverges: it reaches join_ends, then faults while
assigning y_ntv_mod in ntv_mod.f90. This build-dependent endpoint raises
an initialization, lifetime, or ordering concern in addition to possible
join conditioning.

Environment

• branch head: fix/flux-pumping/helcore-grid-parity
• commit: 4b00b887b6a9c37928b6575130b9ad494d2cdc2b
• libneo: 13387c26305fec88158ff109516a10729e2df206
• Debug compiler: GNU Fortran 16.1.1
• MPI: Open MPI 5.0.8, one rank
• related: QL: source_flux reads npl(istep+-2) past bounds -> heap corruption on helical-axis geometries #123, PR QL: preserve full-step endpoints after mode-3 refinement #126

Evidence

The bounds-checked run took 147 s and used about 810 MiB RSS. It had no
npl(istep +/- 2) bounds failure, heap corruption, NaN, repeated zhseqr
failure, or segmentation fault before the join. Three mode-3 retries completed
normally. No final ntv_out, fulltransp, or multispecies HDF5 result was
produced.

[krystophny]

A clean SLURM reproduction at the same NEO-2/libneo heads with GNU Fortran 12.2.0 and the same Debug/bounds flags diverged from the GNU 16.1.1 result:

nl,nr = 24 24
Fortran runtime error: Allocatable argument 'y_ntv_mod' is not allocated
#3 __ntv_mod_MOD_write_ntv_output_a at NEO-2-QL/ntv_mod.f90:813
#4 diag_propagator_res at NEO-2-QL/propagator.f90:953
#5 propagator_solver_loc at NEO-2-QL/propagator.f90:1452

The run passed all three focused source/parity tests, cleared the original #123 bounds failure, completed in 5m37s with about 745 MiB MaxRSS, and reached final joining before the unallocated-output failure. This reinforces the initialization/lifetime/order branch of the diagnosis: compiler/build details change which latent failure is exposed first. The generated qflux files remain unaccepted because the executable exited nonzero before final output completion.

[krystophny]

At descendant 377226441848b64f39e97462937797e10342bab5 (PR130 extension), the same GNU Fortran 16 Debug/bounds case completed successfully:

exit 0; wall 158.32 s; max RSS 809700 KB
nl,nr = 24 24

Four normal mode-3 retries completed. There was no DGBSV diagnostic, bounds failure, NaN, zhseqr, segmentation fault, or backtrace, and final efinal, fulltransp, and ntv_out artifacts were produced. A nontrapping IEEE_INVALID_FLAG summary still appeared after successful completion.

This removes the immediate reproducer but does not establish a join fix: PR130 changes output-state lifetime after joining, while the earlier DGBSV failure occurred during joining. The build/order sensitivity and accumulated IEEE invalid flag still require a replayable join fixture or a trapped phase-local reproduction before #127 can be closed.