Default to the WRF output and rank settings measured in #89

[aethr]

Makes the #89 findings the defaults. Stacked on #90, so it can adjust the rank-count call #90 added for quilting. Measured results are in #89 (comment).

Changes

Setting	Where	Measured effect (aust10km, c8id.16xlarge)	Output
history_begin_h = num_hours_spin_up	setup_for_wrf.py	−16%: 144 fewer frames written	byte-identical
use_netcdf_classic = .true.	setup_for_wrf.py	−33%: 8.7 s → 2.5 s per frame	byte-identical
debug_level = 0 (was 100)	setup_for_wrf.py, and removed from every domains/*/namelist.wrf	−7% of compute	byte-identical
Ranks = physical cores	safe_mpi_tasks(), new physical_cpu_count()	−13% on top of the above	rounding-level differences (see #89)

Together: wrf-run 8,125 s → 4,163 s.

Notes

• Where the settings live: the three output settings are set in setup_for_wrf.py when the namelist is generated, not in the domain namelists, because they describe how we run wrf.exe rather than the domain. history_begin_h follows the config's spin-up length. debug_level = 100 comes out of the domain namelists for the same reason. A run can still override any of these through WRF_NAMELIST_OVERRIDES (Set wrf.exe namelist values and mpirun arguments from the environment #90), which is applied afterwards.
• Physical cores: physical_cpu_count() counts distinct (physical id, core id) pairs in /proc/cpuinfo, the same way the [om-metrics] entrypoint does. It falls back to os.cpu_count() where those fields are missing, as on aarch64, which has no SMT anyway. NCPUS still overrides the default.
• Rank count by host: on the current 64-vCPU WrfStorageSpot hosts the default drops from 64 to 32 ranks. A 96-vCPU host would get 48.
• Classic output: this only affects the history frames wrf.exe writes. average_fields() still compresses the hourly WRFOUT files, and nccopy_compress_output.sh converts any classic files left over at cleanup.

Testing

• make test: 45 passed, including new tests for physical_cpu_count() (hyperthreaded, two-socket, no core information).
• A domain namelist with the three settings applied writes them into &time_control as expected.
• Each setting was run end to end in sandbox through WRF_NAMELIST_OVERRIDES / NCPUS in wrf.exe spends over half its run writing history output; cut that before sizing ranks #89, against an in-batch control. This branch's defaults themselves haven't been run end to end.

🤖 Generated with Claude Code

[aethr]

Closing in favour of #92, which includes the same perf changes but not the complexity of parameterising all the settings.