[willend]

Two fixes around --scan_split, so that split scans match the normal (serial) scan.

1. mctest: --scan_split=auto for %Scan tests only

6253f3f (--statfactor) inserted its block between the %Scan branch and the --scan_split=auto that belongs to it, so:

• with --statfactor (and no MPI), every test got --scan_split=auto: a plain %Example then fails in mcrun with CRITICAL: No interval range specified
• without --statfactor, no %Scan test was run split

Fix: move the two lines back under if test.scan: (as introduced in e5e7b58).

mctest --no-mpi, TAS1_Diff_Slit	runtime errors	%Example split	%Scan split
--statfactor 0.05	0 (before: 1)	no	yes
no --statfactor	0	no	yes (before: no)

2. mcrun: NeXus --scan_split merge as a serial NeXus scan writes it

Each parallel step writes dir/i/mccode.h5 with its data as entry1; these were merged with h5copy (/entry1 → /entryN), keeping references to entry1 and the step dir:

• the target attribute of every linked monitor in entry2.. pointed to /entry1/instrument/.... nexusformat / NeXpy follow it and showed entry1's data for every scan step (h5py-based readers were not affected)
• Directory, name, title, file_name named the step dir dir/i
• the step dirs with their mccode.h5 copies were left behind

merge_nexus_steps() now copies the entries with h5py, rewrites those references (keeping fixed-length strings) and removes the step dirs.

Failed steps: a failed step can leave a partial, unreadable dir/i/mccode.h5, on which the merge (by file existence, with h5copy as well) crashed mcrun before mccode.dat was written. Now only the steps with detector values (the rows of mccode.dat) are merged, numbered entry1..N as a serial NeXus scan does, and all step dirs are kept when a step failed.

middle of 3 steps fails	files	mccode.dat	mccode.h5
NeXus normal (blueprint)	mccode.dat, mccode.h5	2 rows	entry1 (w=0.02), entry2 (w=0.06)
NeXus split, before	0/ 1/ 2/, broken mccode.h5	missing (mcrun crashed)	-
NeXus split, now	mccode.dat, mccode.h5 + kept 0/ 1/ 2/	2 rows	as normal, item by item apart from times

Worker logging: split-scan steps run in multiprocessing workers. With the spawn start method (macOS, Windows) these lacked mcrun's log setup: warnings were printed bare (Scan step 1 failed ..., no WARNING:) and their INFO lines (e.g. call to mpirun failed) were lost. A Pool initializer now sets up mcrun's logging where the worker has none (with fork, Linux, it is inherited, so no duplicate handler), and steps are numbered i+1/N as in the serial scan: WARNING: Scan step 2/3 failed (...). Checked on macOS with spawn and a forced fork: one formatted line per message.

Tested (macOS, mcstas, same seed): 3-point scans of a 0D/1D/2D test instrument and of Test_NeXus_output (+ a dummy scan parameter: 0D, 1D, 2D, Monitor_nD 2D/3D BINS maps, event list), normal vs --scan_split=auto:

• ASCII: identical file layout and data (already before this PR)
• NeXus: same layout (mccode.dat, mccode.h5 only), and mccode.h5 identical item by item (all groups, datasets, attributes incl. types) apart from times; every target resolves to its own entry; nexusformat reads the same data as h5py
• per-step seeds identical in all four cases (1234, 2258, 3282)

Dependency: NeXus --scan_split scans now need h5py in mcrun (instead of the h5copy tool). In mcstas-suite-feedstock h5py currently only comes indirectly (mcstas → nexpy, mcstas-vis → mcstasscript); it should be added to mcstas-core (and likewise for McXtrace).

Not tested on Linux/Windows or with McXtrace (shared code).

🤖 Generated with Claude Code