Allen Coral Atlas SDB (Bahamas/N. Caribbean); multi-source pipeline improvements

[bkeepers]

Summary

Ingests Allen Coral Atlas 10 m satellite-derived bathymetry (SDB) for the Northern
Caribbean / Bahamas as sources/aca_ncaribbean (CC-BY 4.0; raw mirrored to R2), plus several
general multi-source pipeline improvements. The ACA dataset is being tabled — the recipe
runs end-to-end, but SDB isn't good enough for a nautical chart. This PR preserves the work and
documents the challenges for a later revisit. Not intended to merge as-is (keeping it on a
branch leaves main's build untouched).

Pipeline improvements (keep — independently useful)

• tile_raster.py — split an oversized source raster into memory-safe per-tile rasters
  (windowed gdal_translate; drops all-nodata tiles). Also the tool for the deferred GSC Pacific ingest.
• source_smooth.py — gaussian denoise over water for speckly sources; optional --max-depth cutoff.
• source_datum.py — --scale (unit conversion applied before negate; e.g. ACA cm→m).
• source_polygonize.py — generalize large/speckly coverage masks before polygonizing
  (pixel-exact polygons of scattered SDB reefs overflowed the SQLite blob limit).
• aggregation_run.py — removed SKIP_CONTOURS / SKIP_SMOOTH; aggregation stages
  (reproject/merge/smooth/tile/contour) now run independently via a step <name>... CLI instead
  of being skipped by env flags.
• Priority quality tiers — noaa_s102=2, cudem/cudem_third=1 (real surveys outrank
  lower-confidence sources in overlap).

Why ACA is tabled — the challenges

Research confirmed ACA is the only open, dense, redistributable bathymetry for the Bahamas banks
(the Andros "Median DEM" is paywalled; ICESat-2 ATL24 is open but sparse along-track; TCarta/EOMAP
are proprietary). But passive-optical SDB is fundamentally hard for a safety chart:

1. Vertical noise. ±1–2 m with per-pixel speckle. It needed ~σ8 (≈80 m) gaussian smoothing to be
   usable — i.e. its honest resolution is ~z11 (~69 m), not the nominal 10 m.
2. False-shallow over deep water. ACA returns shallow (<18 m) values over water GEBCO shows is
   1,500–4,800 m deep. A self-referential depth cutoff can't catch this (the noise reads shallow).
   A GEBCO-depth-gated feather (w=1 where GEBCO ≥ −12 m, ramp → 0 by −25 m; out = w·ACA + (1−w)·GEBCO)
   does fix the terrain (deep→GEBCO, shallow→ACA) and is a reusable building block.
3. Contour explosion (the blocker). Even after the gate, contouring the vast shallow banks at the
   pipeline's fine 1 m intervals yields ~1 M features per macrotile; tippecanoe takes 45 min+
   to tile that (a global build would be untenable). The only lever that cuts the count is coarsening the
   shallow interval — but fine-grained shallow contours are essential for navigation, so that's not
   acceptable. Neither σ nor the GEBCO gate reduces the feature count.

Revisit path

• Ingest ICESat-2 ATL24 (open, NASA public domain, RMSE ~0.3 m) as sparse depth control / QA,
  and/or to bias-correct ACA.
• Or re-derive SDB with C-SHELPh (open) on ICESat-2 + Sentinel-2 for a cleaner, de-speckled owned grid.
• The GEBCO-gated feather (fixes the terrain) is the reusable piece either way.

See the ROADMAP.md Caribbean table for the tabled entry.