flux-kegonsa: LES flux-footprint corpus and emulator weights, Kegonsa Solar Array tower

Assets for https://github.com/tyatharva/flux-kegonsa (documentation: https://flux-kegonsa.readthedocs.io). Fetch and verify with the repository's bin/fetch_assets.sh; every file's sha256 is in assets/SHA256SUMS there.

path	what
corpus/corpus_cone.h5	1366 (input, target) pairs, 122² @ 30 m grid zero-padded to 128², target cropped to the wind-aligned cone; the training set
corpus/corpus_raw.h5	the same records with the periodic wraparound still in the target
corpus/pairs_npz.tar	the 1366 source records, one .npz per case
corpus/logs.tar	the 8 corpus-machine run logs
seeds/<seed>/seed_restart.nc	the 30-seed FastEddy restart library (2.0 sim-h spin-ups, 122³ @ 30 m)
weights/fno/seed{0..4}/best.pt	the FNO emulator, five seeds
weights/cfm/{seed0..4,long1000_seed0}/best.pt	the conditional flow-matching emulator
predictions/test/	the one audited read of the 2025 test split: FNO predictions and CFM samples
predictions/val/	the FNO's validation predictions

Records: train 837 (2021–2023, 2026) / val 235 (2024) / test 294 (2025). Inputs are the six Kljun et al. (2015) scalars (h, u*, σ_v, L, sin wdir, cos wdir); the receptor is 30 m above ground (28.5 m aerodynamic), a deliberate resolution choice, and there are no stable cases. Generated with NCAR FastEddy v5.0.1 plus the patch series in the repository, forced by HRRR analyses, footprints by the repository's backward LPDM. Licence Apache-2.0.