A bit long report By Claude but confirmed by me that it is true

GMT grdblend: inputs that need reformatting or resampling come back garbled (Windows)

Found 2026-10-08 blending two DGT MDS-2m GeoTIFF tiles (2 m, pixel registered, ETRS89 / PT-TM06):
MDS-2m-196014-04-2024.tiff + MDS-2m-196015-04-2024.tiff, -R-4000/-3000/-287000/-285000 -I2.

Symptoms

Also printed: Failed to remove c:\TMP/grdblend_reformatted_XXXX.nc! [remove error: Permission denied].

Cause (grdblend.c, grdblend_init_blend_job)

1. Reformat branch (non-netCDF input -> grdconvert):
   gmt_get_tempname(GMT->parent, "grdblend_reformatted", ".nc", buffer) uses ONE fixed template.
   With MSVC _mktemp_s returning the same name every time (the comment in the resample branch
   says so), every reformatted input is written into the same file: the second overwrites the
   first while it is still open (hence the failed delete). The resample branch already works
   around this with sprintf(template, "grdblend_resampled_%d", n); the reformat branch does not.
   Probable fix: the same per-input template, e.g. "grdblend_reformatted_%d".
2. After either branch the re-read grid goes through
   grdblend_overlap_check (GMT, &B[n], h, 0) UNCONDITIONALLY. That is the longitude +-360 check
   (its earlier calls are guarded by gmt_M_x_is_lon); on a Cartesian/projected grid it "adjusts"
   x by 360 (-Vi: "grid region needed longitude adjustment", -Vd: out: -360/140), so only
   140 of the tile's 500 columns land on the output. This is the main cause of the wrong output.

MWE (Julia, GMT.jl)

using GMT
a = raw"C:\Users\j\.gmt\DGT\MDS-2m\MDS-2m-196014-04-2024.tiff"
b = raw"C:\Users\j\.gmt\DGT\MDS-2m\MDS-2m-196015-04-2024.tiff"
R = grdblend("\"$a\" \"$b\""; R="-4000/-3000/-287000/-285000", I=2, r=true)
count(!isnan, R.z)            # 140000, expected 445196