FIX spectral indices to use band values as read, with radiometric scaling

[mkiani12]

Stacked on #51 — review #50 and #51 first; this PR targets hotfix/afri-formula.

Problem

All six index calculators read from FileHandler.get_normalized_bands():

(image - np.min(image)) / (np.max(image) - np.min(image))

applied independently to each band. Each band therefore receives a different affine transform, which changes the relationships between bands — and those relationships are the entire physical content of a spectral index. (NIR' - Red') / (NIR' + Red') where NIR' and Red' were rescaled by different factors is simply not NDVI.

Reproduced on example/data/:

	before	after
NDVI mean	0.3351	0.5103
pixels NDVI > 0.3	55.64%	59.88%
max abs difference vs rasterio reference	0.7453	0.0
same pixel, full scene vs 400×400 crop	0.8373 / 0.8692	identical
AFRI vs NDVI correlation	−0.6927	+0.7117
BSI vs NDVI correlation	+0.3908	−0.6507

The last two are the ones I find most telling. Karnieli et al. report AFRI and NDVI as "almost identical" under clear sky, and bare ground must oppose canopy. Both relationships held on the stored values and inverted under the rescale — the index was not merely rescaled, it was reporting the opposite of the physical truth. Those are the two indices whose formulas #51 just corrected; they only start behaving correctly here.

The extent dependence is the reproducibility failure: loading a crop changed the answer for pixels the crop shared with the full scene.

Changes

1. Indices compute on values as read. New FileHandler.get_bands() returns the arrays already held in self.bands, mirroring the existing accessor pattern. The six calculators switch to it; self.normalized_bands becomes self.source_bands.

Min–max normalization is untouched for the enhancement, HSV, PCA and SVM modules, where rescaling is appropriate — and for SVM it is genuinely required, since an RBF kernel needs comparable feature scales.

2. Radiometric scaling, because removing the rescale is not sufficient on its own. It leaves the indices computing on digital numbers, and two of them carry constants defined in reflectance units:

Index	Constant	Safe on raw DN?
NDVI, NDWI, UI, BSI	none	Yes — invariant to a gain applied to all bands equally
SAVI	soil adjustment L = 0.5	No
AFRI	coefficients 0.66 / 0.50 on SWIR	No

Adding L = 0.5 to a DN in the thousands contributes nothing, so SAVI on unscaled input is silently not SAVI. Every index now accepts scale_factor/offset, applied as ρ = DN × s + o, with published values for the common analysis-ready products shipped as RADIOMETRIC_PRESETS:

SAVICalculator(nir_path=..., red_path=..., **RADIOMETRIC_PRESETS["landsat-c2-l2"])

Covering Landsat C2 L2 (2.75e-5, −0.2), Sentinel-2 L2A (1e-4, 0), and Sentinel-2 L2A baseline ≥ 04.00 (1e-4, −0.1, for the BOA_ADD_OFFSET introduced in that baseline). SAVI and AFRI emit a UserWarning when handed values far outside the reflectance range, so the silent-nonsense case becomes loud.

Behaviour change

Index values change for every user — deliberately, and documented prominently in docs/spectral-indices.md. The previous values were scene-dependent and not the cited indices, so preserving them would preserve the defect. Defaults are scale_factor=1.0, offset=0.0, so input already in reflectance needs nothing passed.

Tests

347 pass (334 on #51, 13 new). tests/tools/spectral_indices/radiometry_test.py covers:

• Extent invariance — a pixel's NDVI is identical for the full array and a crop containing it, plus a companion test asserting the same comparison fails under min–max, so the first is not vacuous.
• Gain invariance — NDVI unchanged when all bands are multiplied by 10000, the property that makes normalized differences safe on DN.
• Physical relationships on the bundled Landsat subset — AFRI must track NDVI (> 0.5) and BSI must oppose it (< 0), with both asserted to invert under per-band rescaling. These read the real bands via rasterio and skip if unavailable; synthetic noise has no spectral structure to preserve, so this property can only be tested on genuinely correlated bands. My first attempt used random arrays and was measuring nothing.
• Scaling arithmetic, calculator-level application, and the SAVI/AFRI warnings firing on DN while NDVI stays silent.

The six existing index test fixtures were updated from get_normalized_bands/normalized_bands to get_bands/source_bands.