[bpurinton]

DEMBenchmark.run(pc_align=True) already reuses the pc_align products under its folder, so a re-run makes no pc_align call. It still rewrote each DEM's translated copy (dem_benchmark/<label>/<stem>_pc_align_translated.tif) on every run, a full-size raster write per DEM. Re-scoring the 39 full-scene DEMs of #207 from cached alignments spent about 6 of its 20 minutes on those writes.

Change

• The translated copy is reused when its mtime is at least that of the DEM and of every pc_align/<prefix>-* file (_is_newer_than). Otherwise it is rewritten.
• A rewrite goes to *_pc_align_translated.partial.tif and is then moved into place with os.replace, so an interrupted run never leaves a partial copy that the next run would take as current.
• The docstrings of the class's directory argument and of run(pc_align=...) say so.
• One [Unreleased] / Changed line in the CHANGELOG.

Tests

TestRun now also checks three things:

• a second run leaves the translated DEM's mtime unchanged and returns the same table;
• a translated DEM older than its pc_align output is rewritten;
• no .partial file is left behind.

pytest tests/test_dem_benchmark.py: 34 passed. pre-commit run --all-files: passed.

Commits

1. A cherry-pick (-x) of ac1db60 from Full-scene scene-combination benchmarks: Atlanta and UCSD, lidar check, Atlanta jitter (#207) #210, where the change was first made.
2. A rewording of the CHANGELOG line, which had quoted that study's numbers.

#210 itself will be closed: its notebooks and write-up are moving out of this repo.