Phase 1 of #36: the offline, network-free core of bounding dft_stac_fetch()'s
download to the AOI footprint. tile_size_check() validates a tile_size and snaps
it to a multiple of res (so tile pixel grids align for a seam-free merge);
tile_grid() splits the AOI bbox into a res-aligned grid anchored at the lower-
left and keeps only tiles intersecting the AOI polygon — a thin corridor fetches
near its footprint, not its full bbox. Boundary tiles are left un-trimmed (the
overhang is masked away; trimming would break res-alignment). Tests written
first (validation, snapping, intersecting-subset, alignment, single-tile,
degenerate-AOI); 40 pass / 1 skip; lint clean; code-check clean.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DKqnW6Fni3tc44Lo5D5fJz

Phase 2 of #36: stac_cache_key() gains an optional tile_size, appended to the
hash only when non-NULL. An untiled fetch (tile_size = NULL) reproduces the
exact pre-tiling 9-element hash, so existing io-lulc caches stay valid on
upgrade; a tiled fetch keys distinctly, so its terra .tif mosaic is never served
as an untiled gdalcubes .nc (or vice versa). A frozen golden-hash regression
guards the legacy-preservation invariant. tile_size arrives already snapped by
the caller (single normalization site — no gate/key desync). 45 pass / 1 skip;
lint + code-check clean.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DKqnW6Fni3tc44Lo5D5fJz

#36)

Phase 3 of #36: pull the cube_view + raster_cube + write_ncdf block out of the
per-year loop into fetch_extent_to(). The untiled path now routes through it,
writing straight to the existing <yr>_<key>.nc cache — identical filename,
format, and I/O to before (faithful code-motion; cube_view built identically).
Sharing this primitive is what will guarantee each tile fetches identically to
the corresponding slice of the untiled cube when the tiled branch lands in
Phase 4. 45 pass / 1 skip; lint clean.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DKqnW6Fni3tc44Lo5D5fJz

Phase 4 of #36: tiled fetch end-to-end. When tile_size is set, dft_stac_fetch()
splits the AOI bbox into a res-aligned grid, streams only tiles that intersect
the AOI polygon (via the shared fetch_extent_to()), mosaics them with
terra::merge(terra::sprc(...)) (new mosaic_tiles() helper), and caches a terra
GeoTIFF — so a thin corridor fetches near its footprint instead of its full
bounding box. Extension routes on is.null(tile_size) (.nc untiled / .tif tiled),
matching the distinct cache key. tile_size is normalized once (snap to res
multiple) so the path gate and cache key agree; GDAL /vsicurl config (mirroring
dft_stac_cube) is set with an on.exit restore, scoped to the tiled path, since
tiling multiplies per-item COG opens. Default tile_size = NULL is unchanged.

Offline merge oracle (byte-for-byte reassembly, single-tile, .tif round-trip)
and an opt-in DRIFT_TEST_NETWORK e2e (tiled == untiled over the in-AOI overlap,
extension routing). 352 pass / 5 skip; document + lint + code-check clean.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DKqnW6Fni3tc44Lo5D5fJz

Phase 5 of #36. Document the download-tiling workaround in the gdalcubes gotchas
note (the filter_geom-independent way to bound the read, cross-referencing the
cube-path residual #38); add the NEWS 0.6.0 entry for tile_size; bump DESCRIPTION
0.5.0 -> 0.6.0. devtools::check() clean: 0 errors / 0 warnings / 0 notes.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DKqnW6Fni3tc44Lo5D5fJz

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01DKqnW6Fni3tc44Lo5D5fJz