Add a `co3d-cars` reader (CARS tiled LAZ) and run the CO3D -> COPC arm

[emmanuelmathot]

Motivation

CO3D is the only scoped mission with no measurement at all. The mission set was locked at four (Sentinel-1, Sentinel-2, SWOT, CO3D) and the other three are measured, so the CNES report (D2 ch. 6.6) still states its COPC recommendation as provisional and its summary-table row reads pending. That is the last mission-shaped hole in the final benchmark campaign.

The harness side is mostly done already: the COPC writer, the point-cloud read path (bbox query over octree nodes) and the octree level-of-detail figure all landed with SWOT PIXC (#28, #38, #42, #44). What is missing is a reader for the CARS delivery shape, plus the sample.

Ask

Add a co3d-cars dataset reader and run the arm.

• Reader. CARS pipeline output is delivered as tiled LAZ, one file per tile, so the product is a set of tiles, not a single granule. Enumerate the tiles under source (*.laz), one component per tile, and keep the tile identity in the component name so the object-size distribution is per tile.
• Arm. co3d-cars to COPC, reusing the PIXC adapter path. Metrics: write, object_size, read (bbox query). Report the per-tile object-size distribution against both Datalake tiers, and the octree depth.
• Grouping lever. The tile is the delivered unit and may sit below Tier 2, so record whether the lever here is the octree node budget inside one tile or merging tiles into one COPC (see the aggregation issue in this batch).

Acceptance

• co3d-cars registered in DATASETS, unit-tested on a synthetic LAZ tile set.
• An example benchmark config (configs/benchmarks/example_co3d_cars_copc.yaml) + dataset config that validate.
• A committed run once the sample lands, with the per-tile size distribution, tier verdict and octree depth.

Blocked on the sample, not the harness

The T-Rex path is still a placeholder in the CNES testbed (testbed/sources/co3d-cars.yaml points at ..._DATASET/CO3D_CARS_TBC) and no open-data sample is confirmed. Confirming the path and staging a representative tile set is the CNES-side prerequisite; the reader can be written and tested against synthetic LAZ before that.

Touch points

src/cng_benchmark/datasets/ (new co3d.py), configs/datasets/, configs/benchmarks/; reuses formats/copc.py.

Related

• CNES study: D2 ch. 6.6, consolidated-metrics.md gap 3, last-mile action 18.
• Point-cloud path this reuses: SWOT-C: swot-pixc point-cloud reader (PDAL) + COPC adapter (PIXC -> COPC) #28 (PIXC reader + COPC adapter), feat(copc): carry the full PIXC pixel_cloud variable set (content-complete COPC) #38 (content-complete), feat(copc): publish the octree level-of-detail figure as a systematic run output #44 (LoD figure).

[emmanuelmathot]

Deferred: the sample was located in the Datalake cnes-cng-study bucket (falaises_LETORTU/) but the credential that reads it is not available on our side, and the runner's sprid role is refused on that bucket (probe: sprid list OK in 0.12 s, this bucket AccessDenied in 0.04 s, so authorisation rather than reach). The reader is written and validated on a synthetic fixture in #89. The CNES study report now states CO3D as not measured, with its COPC recommendation transferred from the measured SWOT PIXC path. To resume: a credential on that bucket, then one run, no development.