surtgis v0.14.7

Patch release. Advances ROADMAP item M ("Pansharpening") from
the 2026-06-08 OTB-survey backlog. Ships three classical
pansharpening algorithms covering the practical spread: a
spectral-ratio method (Brovey), a covariance-based method (PCA),
and a Gram-Schmidt orthogonalisation method (Laben & Brower 2000).
All operate on arbitrary band counts and follow a common contract.

Added

• pansharpening::brovey — Brovey transform (Gillespie et al.
  1987). out_b = MS_b · pan / mean(MS). Fast, simple, can
  oversaturate radiometric values when synthetic and real pan
  differ strongly.

• pansharpening::pca_pansharpen — PCA pansharpening (Chavez
  et al. 1991). Builds the covariance matrix of the MS bands,
  diagonalises it with the shared Jacobi solver from
  classification::pca::jacobi_eigen (now pub(crate)),
  histogram-matches the pan to the first PC, and back-projects.
  Works for any number of bands.

• pansharpening::gram_schmidt — Gram-Schmidt pansharpening
  (Laben & Brower 2000, US patent 6,011,875 expired January 2018,
  now public domain). Builds the centred Gram-Schmidt sequence
  (P_lr, MS_1, ..., MS_B), records the regression coefficients
  φ_{k,i}, substitutes the histogram-matched pan into the first
  GS slot, and inverts. The same φ coefficients are used in
  the forward and inverse passes — that's the elegant property
  that makes the substitution clean.

• CLI: surtgis imagery pansharpen with three subcommands:

  • brovey --pan PATH --band PATH [--band ...] --output-dir DIR
  • pca --pan PATH --band PATH [--band ...] --output-dir DIR
  • gram-schmidt --pan PATH --band PATH [--band ...] --output-dir DIR
    Each writes one <prefix>_band<NN>.tif per output band into
    output-dir.

Contract

All three functions take pan: &Raster<f64> and ms: &[&Raster<f64>]
with the MS bands already upsampled to pan resolution (upstream
caller's responsibility — use vector::resample or any other
bilinear/cubic upsampler). They return Vec<Raster<f64>> of the
same length as ms. Pixels where any input is NaN propagate to
NaN in every output band; output nodata is set to Some(NaN).

Internal change

classification::pca::jacobi_eigen is now pub(crate). PCA
pansharpening reuses it instead of duplicating the symmetric-matrix
eigensolver — single source of truth for the numerical kernel.
Public API of classification::pca is unchanged.

Tests

• Brovey: pan-equal-synth preserves MS bit-for-bit, pan-doubled
  doubles every output, NaN propagation, zero-synth guard,
  empty-bands and mismatched-shape errors.

• PCA: pan-matching-PC1 yields near-lossless reconstruction
  (≤ 5 units mean-abs drift on a 4×4 gradient test), NaN
  propagation, empty/mismatched-shape errors.

• Gram-Schmidt: pan-equal-synth recovers MS at 1e-9 tolerance
  (the deepest correctness check — pins the orthogonalisation +
  inversion math), per-band mean preservation property
  (centred-pan injection has zero mean by construction), NaN
  propagation, empty/mismatched-shape errors.

15/15 pansharpening tests pass.

Backlog status

Survey 2026-06-06 status after this release: G ✓, H ✓, I
trigger-driven, J ✓ (v0.14.4), K ✓ (v0.14.5), L ✓ (v0.14.6),
M ✓ (this release). Remaining: N (MAD / IR-MAD), O (mosaic
seams + color balance).