Bound dft_stac_cube() streaming to the AOI (fetch-time read still bbox-bound after #32)

[NewGraphEnvironment]

Problem

dft_stac_cube() builds its gdalcubes cube over the AOI's bounding box (cube_view(extent = bbox_target)), so the COG streaming — the dominant cost, ~10–30 min for a multi-year monthly Sentinel-2 fetch — scales with the bbox, not the AOI. For a thin, corridor-shaped floodplain the bounding box is largely empty.

#32 restored polygon-tight output (clip = TRUE masks the assembled stack with terra::mask()), but that runs after the pixels are streamed — it does not reduce what gdalcubes reads, so fetch time is unchanged.

Measured on the packaged example AOI (Neexdzii Kwa reach): area / bbox ≈ 0.105, i.e. ~90% of the streamed bounding box is outside the polygon — a ~10× streaming overhead. Thinner / more diagonal reaches are worse (NECR io-lulc: ~10.1% inside, per #36).

Distinct from #32

#32 clips the output of dft_stac_cube() (via terra::mask(), since gdalcubes::filter_geom() segfaults on the pinned build — see inst/notes/gdalcubes-pc-gotchas.md). This issue is about reducing the read: bounding the COG streaming to near the AOI footprint.

Related to #36

#36 is the same download overhead on the categorical dft_stac_fetch() (io-lulc) path, with a tiling proposal (fetch only tiles intersecting the AOI, then mosaic). The technique transfers, but this issue tracks it for dft_stac_cube() (continuous Sentinel-2), which has its own surface: a per-tile cube_view, cloud/SCL masking, the baseline-offset item split at 2022-01-25, and the terra::cover() coalesce.

Proposed

Options, roughly in order of feasibility:

• Tile the cube_view — split the AOI bbox into a grid, build/stream one cube per tile intersecting the AOI, skip empty tiles, and mosaic (mirrors Tile dft_stac_fetch to bound download over sparse-floodplain bounding boxes #36). Peak read approaches the union of intersecting tile bboxes — near the AOI footprint for a corridor.
• A working filter_geom — if/when the gdalcubes segfault is fixed upstream, it pushes the AOI into the read directly (no output-side terra::mask() needed either).
• Server-side windowing — any STAC/gdalcubes path that reads only the AOI window.

Related

• dft_stac_cube: restore AOI-polygon clip (filter_geom segfault workaround) #32 — output clip restored via terra::mask() (this is the read-side residual)
• Tile dft_stac_fetch to bound download over sparse-floodplain bounding boxes #36 — same overhead on dft_stac_fetch(); shared tiling technique
• inst/notes/gdalcubes-pc-gotchas.md — records the filter_geom segfault, the dft_stac_cube: restore AOI-polygon clip (filter_geom segfault workaround) #32 resolution, and this residual