Floodplain-wide tiled composite: terra::mask cannot read the merged mosaic after a 3-hour read

[NewGraphEnvironment]

Problem

A floodplain-wide dft_stac_composite() over BULK (bulk_co_ff04, tile_size = 20000, 2023 Jul–Aug, clip = TRUE) reads every tile and then fails after the read, on the merged mosaic:

Error: [mask] cannot read from /private/tmp/Rtmp.../spat_16d103f4b1e92_93456_SskviV5XwhZb4QV.tif

That is stac_cube_clip() → terra::mask() failing to read the temp GeoTIFF that mosaic_stacks() (terra::merge(terra::sprc(tiles))) wrote. The same path is dft_stac_cube()'s tiled read, so a floodplain-scale tiled cube is presumably affected too.

Established (drift#79 BULK runs, 2026-09-28)

run	what happened
2	30 tiles in 54 min, 15 of them lost to expired tokens (fixed in #79). Died at writeRaster(filetype = "COG"), which reads the same merged mosaic
4	all 30 tiles read, 0 chunk failures, in 3 h 16 min; peak RSS 3.90 GiB; then the mask error above

The R session exited on the error, so its tempdir, and the evidence, went with it.

Ruled out

• Disk: 1 TB free on /private/tmp.
• Writing a floodplain-size COG: an in-memory synthetic raster with the same dims and options writes fine (82 s, 459 MB).
• terra deleting a temp file when its parent object is collected: subset, reassign, gc() and mask() on a disk-forced merge all work (terra 1.9.50).
• Merge plus mask at this scale, in general: 30 synthetic NetCDF tiles on the exact BULK tile grid, merged to a file-backed spat_*.tif, reordered by name and masked with the real 104,584-vertex floodplain polygon. OK in 0.7 min.

Not yet separated

What the live run has and the offline rebuild does not:

• gdalcubes' float64 NetCDF with a time dimension as the merge sources;
• 3 hours of elapsed time, and temp files created early being read late;
• per-tile NA patterns from cloud masking.

data-raw/benchmark_composite_bulk.R now catches the error and inventories the temp files before exit: names, sizes, mtimes, and gdalinfo -checksum of every spat_*.tif. The next floodplain run leaves the evidence this one lost. A cheaper way to separate the first candidate is the offline rebuild with tiles written by gdalcubes::write_ncdf() from a local image collection. The #79 review built such a collection in its scratch scripts.

Acceptance

• A floodplain-wide BULK composite completes and caches one COG.
• A test pins the cause (fixture from real gdalcubes NetCDF tiles if that is the discriminator).
• The numbers go in the scale-test record: wall time and peak RSS for the full run.

Relates: drift#79, drift#87