Windows CI: a 1-row netCDF grid is read back with a collapsed y range (test/grdtrend/trend_xy_plot.sh) #9261

I asked Claude to look for a solution to the failure of this test in the Windows CI, and this is its analysis. I am posting it to ask what you prefer to do about it.

test/grdtrend/trend_xy_plot.sh fails only in the Windows CI (MSVC 14.51.36231, Visual Studio 2026 image, Release /O2 /Ob2 /DNDEBUG, vcpkg netCDF 4.9.3): grdtrend prints Grid y increment <= 0.0 [data.grd] and crashes. I would like to ask what you prefer for this one, because I could not find the root cause.

What fails. The grid written by grdmath is valid and reads fine on Linux and through GDAL, but on the Windows CI the same call reads a 6x1 grid with y range 0/0:

gmt grdmath -R-1.2/1.2/-0.2/0.2 -I0.4 -r 0.5 X ADD X X MUL ADD X 3 POW ADD = row.grd
gmt grdinfo -C row.grd

• good: row.grd -1.2 1.2 -0.2 0.2 -0.5 3.5 0.4 0.4 6 1 1 0
• CI: row.grd -1.2 1.2 0 0 -0.5 3.5 0.4 0 6 0 1 0 (y range 0/0, y_inc 0, 0 rows)

It fails when the test runs alone (no -j4), and also outside the test (https://github.com/GenericMappingTools/gmt/actions/runs/38145255398). A 1-column x 6-row grid reads fine, GDAL reads y#actual_range={-0.2,0.2} fine, and the only netcdf.dll copies on PATH are vcpkg's. For a 1-row grid the y range comes only from actual_range (there is no usable y vector), in the y block of gmtnc_grd_info (src/gmt_nc.c, around lines 774-803); the x block has the same structure.

What I tried.

• Any debug print inside gmtnc_grd_info, even a passive one, makes the failure disappear.
• #pragma optimize("", off) around gmtnc_grd_info only, with nothing else changed, makes the test pass (https://github.com/GenericMappingTools/gmt/actions/runs/38176185877).
• ASAN and UBSAN builds on Linux find nothing in this path, and I found no undefined behaviour by reading it.
• A gmt 6.8.0_f6039ec build from September on a Windows 10 PC reads the grid correctly, but I don't know its toolset or flags. gmtnc_grd_info is identical to the one in master.

So it looks like an optimizer problem (compiler bug or an undefined behaviour that /O2 exploits), but I could not prove either.

Questions.

1. Does anyone with a local MSVC 14.50 or 14.51 build reproduce the grdinfo output above?
2. The 6.7.0 Windows binaries were linked with MSVC 14.50. Does the official 6.7.0 binary read the 1-row grid correctly? If not, users are affected and not only the test.
3. How do you prefer to handle it: (a) the #pragma optimize change above, guarded by _MSC_VER (I have a 6-line patch ready), (b) mark the test with # GMT_KNOWN_FAILURE_WINDOWS, or (c) something else?

Reviewed with Claude Opus 5.5 (read-only review of the analysis). Written with Claude Sonnet 5.5.