edaphos 1.3.1 \u2014 honest Cerrado benchmark repair

Summary

v1.3.0 shipped the first real-data benchmark but four load-bearing
defects made the numbers undersell the stack (R² = 0.24 for B1, E
worse than B1). v1.3.1 repairs each defect transparently, with the
repaired numbers defensible against the published literature.

What changed

• Target: from "any horizon with lower_depth ≤ 30" (mixing
  0–5 / 5–15 / 15–30 cm slices) to "shallowest surface-anchored
  horizon (upper_depth == 0, lower_depth ∈ [5, 30] cm)" per
  profile. One physical quantity across every row.
• Positional uncertainty: relaxed from ≤ 500 m to ≤ 2 km,
  matching the 1 km covariate resolution.
• Covariates: added ESA WorldCover 2020 fractional covers
  (Zanaga et al. 2021) and 19 WorldClim 2.1 bioclim indices (Fick
  and Hijmans 2017). 32 → 56 covariates.
• Evaluation: replaced the single 80/20 split with
  5-fold spatial cross-validation (k-means on coordinates).
  Every profile is a held-out prediction exactly once.
• Tried and abandoned honestly: an integrated 0–30 cm SOC
  stock target. WoSIS's per-horizon bulk density covers only
  ~20 % of Brazilian profiles; the stock formulation degenerated
  into a constant-BD-fallback target with weaker signal than the
  plain concentration. Documented in the vignette and NEWS.

Headline numbers (5-fold CV, 1095 profiles)

Method	n	RMSE (g/kg)	R²	PICP @ 95	Interval score
B1 ranger QRF	1095	13.51	0.219	0.944	65.8
B2 ranger + gstat kriging	910	13.86	0.233	0.817	99.5
E ranger + MoCo v1 embed	923	14.07	0.157	0.940	71.7

R² 0.22-0.23 is in line with published Brazilian Cerrado DSM
(Gomes et al. 2019: 0.13 Brazil-wide; Nakhavali et al. 2018: 0.28
savanna with 200 profiles). The plain QRF is the calibration
champion (PICP 0.944 at a 0.95 nominal level, best interval
score).

Still pending

The foundation-model embedding (encoder v1, 20 k InfoNCE steps)
trails B1. Encoder v2 with 200 k steps (10× budget) is
currently in training on Apple M1 Max MPS; v1.3.2 will re-run the
benchmark with the v2 weights and publish a new Zenodo deposit.

Scripts

• data-raw/case_cerrado_prepare.R + _run.R — fully reproducible
  pipeline (~1 h end-to-end).
• data-raw/pretrain_cerrado_train_v2.R — reproducible 200 k step
  encoder retrain.

Quality

• Vignette case-cerrado-end-to-end rebuilt from scratch against
  the 1095-profile CV bundle.
• R CMD check --as-cran: 0 errors / 0 warnings / 1 harmless
  NOTE.