vectorization refactor

[mdsumner]

I'm in the middle of this

compute_spatial_window <- function(locations) {
  locations |>
    mutate(
      crs = mk_utm_crs(lon, lat),
      utm_x = mk_utm_centre_x(lon, lat),
      utm_y = mk_utm_centre_y(lon, lat),
      utm_xmin = utm_x - radiusx,
      utm_xmax = utm_x + radiusx,
      utm_ymin = utm_y - radiusy,
      utm_ymax = utm_y + radiusy
    ) |>
    # buffer_extent and reproj_extent need rowwise or vectorized versions
    rowwise() |>
    mutate(
      utm_extent = list(buffer_extent(c(utm_xmin, utm_xmax, utm_ymin, utm_ymax), resolution)),
      ll_extent = list(reproj_extent_safe(utm_extent, crs))
    ) |>
    ungroup() |>
    mutate(
      lonmin = map_dbl(ll_extent, 1),
      lonmax = map_dbl(ll_extent, 2),
      latmin = map_dbl(ll_extent, 3),
      latmax = map_dbl(ll_extent, 4)
    ) |>
    select(-utm_extent, -ll_extent)
}

Or write thin vectorized wrappers for buffer_extent / reproj_extent that take column vectors and return data frames:

rbuffer_extent_v <- function(xmin, xmax, ymin, ymax, res) {
  pmap_dfr(list(xmin, xmax, ymin, ymax, res), \(xn, xx, yn, yx, r) {
    e <- buffer_extent(c(xn, xx, yn, yx), r)
    tibble(utm_xmin = e[1], utm_xmax = e[2], utm_ymin = e[3], utm_ymax = e[4])
  })
}

might not even need iteration = "list" anymore:

tar_target(spatial_window, 
           compute_spatial_window(locations), 
           pattern = map(locations)),
# spatial_window: 1 or 2 rows per location (AM split)

tar_target(query_specs,
           prepare_query(spatial_window, collection, provider),
           pattern = map(spatial_window)),
# query_specs: data frame, 1 row per spatial_window row

tar_target(assets_table, 
           get_assets_from_urls(query_specs),
           pattern = map(query_specs))
# assets_table: data frame, auto row-binds

Data frames in, data frames out, targets just binds them together. No list wrangling.
The mental model is now just: locations → (explode at AM) → queries → assets → (regroup by location_id later if needed).