Add a round-trip integrity / lossless-verification arm (sample source vs converted)

[emmanuelmathot]

Motivation

The harness measures size, write, read and display, but asserts nothing about whether a conversion is lossless. For every arm (netCDF → COPC, netCDF → GeoZarr, GeoTIFF → COG, shapefile → GeoParquet) we should be able to prove the converted object actually carries the source's data. A CNES reviewer asked us to verify exactly this for the PIXC → COPC path ("verify we really have everything") at the 2026-06-24 check-in.

Proposed change

An optional integrity metric that, after the target is written, samples values from the source and the target at corresponding locations and asserts equality per variable within a tolerance. Report, per variable: checked yes/no, points/cells/features compared, max abs diff, max rel diff, pass/fail.

• Sampling configurable: samples.integrity = N (random points/cells/features) or all.
• Per-format comparator behind a small interface:
  • point cloud (COPC): match by point identity (index after the finite-mask, or nearest-coord), compare x/y/z + every extra dimension.
  • raster (COG / GeoZarr): compare a set of pixel windows against the source bands/subdatasets.
  • vector (GeoParquet): compare feature geometry + attributes by feature id.
• Surfaces in result.json next to the other metrics; a human line in summary.md.

Acceptance

• integrity appears in result.json with per-variable diffs.
• CI exercises it on the synthetic fixture (lossless ⇒ pass), and a deliberately lossy path (e.g. dropping a variable) ⇒ fail.
• Documented in docs/configuration.md (metrics) and docs/architecture.md (comparator seam).

Related

• Directly verifies the content-complete COPC in PIXC → COPC stores geometry only; carry the full pixel_cloud variable set (content-complete COPC) #36 — this is how we prove the full PIXC pixel_cloud variable set survived the conversion.
• Keeps to the project rule that processing benchmarks stay out of scope; this is a correctness check, not a processing metric.