fix: exact geos inverse + red-band relaxation for limb-correct true color

[LeonJian]

Summary

Fixes the insufficient edge contrast in the true_color_reproduction output image (bottom limb mean ~50/255, flat ~11-22 std) by replacing the flat tangent-plane geos projection inverse with the exact ellipsoidal ray–ellipsoid intersection, and completing the Satpy/pyspectral parity details for the Rayleigh correction.

Root cause

rusty_sat_modifiers computed per-pixel lon/lat with a flat-Earth approximation (lat = atan(tan θ_y · cos θ_x), lon = λ₀ + θ_x), exact only at the sub-satellite point. At the AHI full-disk limb (scan angle ~8.4°) it underestimates ground offsets by the h/a curvature factor (~5.6×), giving latitudes/longitudes off by ~57°. Consequences for the true-color pipeline:

pixel	true sza / corr	old (flat) sza / corr
bottom limb	84.5° / 10.5	29.4° / 1.15
left limb	76.1° / 4.2	26.8° / 1.12
top limb	46.7° / 1.46	15.0° / 1.04

The limb was not sun-zenith-corrected and the Rayleigh LUT path lengths were ~9× too small → dark, flat edges.

Changes

• geos.rs: exact ellipsoidal geos inverse — ray–ellipsoid intersection (quadratic; A separable, B/C constants) + geodetic-latitude Newton iteration; matches PROJ geos / pyresample get_lonlats, verified to 1e-9 against limb reference values (horizon ±81.28°).
• angles.rs: rad-based sun_angles_from_lonlat / sat_angles_from_lonlat; AngleSet and the strip loops route through the exact inverse; deleted precompute_columns/rows + solar/satellite_single_precomputed.
• sun_zenith.rs / rayleigh.rs: strip loops use the exact inverse (space pixels → NaN via discriminant).
• Rayleigh pyspectral parity:
  • LUT boundary: clamp sun/sat zenith angles to the LUT max secant before computing 1/cos (_clip_angles_inside_coordinate_range) — limb/night pixels evaluate at the LUT edge instead of extrapolating past the last grid point.
  • Correction order: interp → relax(red) → clip [0,100] → reduce → vis − corr (pyspectral np.clip semantics; dropped the result max(0)).
  • Red-band cloud relaxation: new RedBandSource (None / Dataset / SunZenithCorrectedVis), wiring Satpy rayleigh_corrected semantics (sun-zenith-corrected B03 as the red band; B03 itself uses the in-place post-sunz values with no extra memory).
• true_color_reproduction.rs: wires the B03 red band (0.5 km self + 1 km nearest-resampled via the existing pipeline), replaces the pre-fix artifact assertion (Rayleigh monotonicity does not hold under sunz amplification), and adds edge-contrast regression assertions on the in-memory u8 image (bottom/left limb means > 120, left-limb std > 25, top std > 20, center mean/std unchanged).

Measured impact (true_color_05km.png)

region	old mean	old std	new mean	new std
center	112/116/124	54/51/46	111/115/123	55/52/46
bottom limb	47/54/57	18/19/22	169/179/201	22/17/16
left limb	76/87/95	19/16/14	170/181/206	47/41/26
top limb	129/143/151	21/14/12	143/152/162	32/26/23

Tests

• geos parity table (limb pixels, 1e-9), spherical/ellipsoidal horizon checks
• limb sunz/correction-factor tests (curved-Earth references)
• LUT secant clamping, correction clip order, red relaxation (incl. red > 100 negative factor parity, SunZenithCorrectedVis path gating)
• full workspace: 660/660 tests pass (release), cargo fmt + cargo clippy -D warnings clean

[LeonJian]

Follow-up: scientific corrections, JMA TCR color parity, and performance (3 new commits)

This PR was opened with the exact-geos-inverse fix (5c7a36c). The branch has since grown three commits that close the remaining scientific gaps found by review, add the missing JMA True Color Reproduction enhancement, and cut the integration-test runtime by ~2.1× while keeping the output bytes identical. All work was verified against the inspected Satpy/pyspectral/trollimage reference code.

1. 3585c4e — SZA gradient fix, out-of-LUT sun-zenith fix, JMA day/night blend

Root cause A (whole-image red cast): the 88°–95° sun-zenith falloff was inverted. sza_correction_factor used a cosine-domain gradient with reversed endpoints: 0 at 88° (should be ~28.65, continuous with the 1/cos branch) and 28.65 at 95° (should be 0). Replaced with the exact Satpy _sunzen_corr_cos_ndarray angle-domain formula; reference checks now lock 88°/90°/94.5° → 28.65/18.26/1.50.

Root cause B (edges redder than the interior): the combined correction skipped the sun-zenith amplification for wavelengths outside the Rayleigh LUT (400–800 nm). AHI B04 (0.86 µm) — the 15% NIR term of hybrid_green — was returned raw, so the green channel was too dark at high SZA and the red cast grew toward the limb. The combined path now applies the sunz amplification unconditionally and the Rayleigh subtraction only for in-LUT wavelengths (Satpy [sunz_corrected]-only semantics for B04).

Also in this commit:

• daynight_blend_weights (Satpy DayNightCompositor._get_coszen_blending_weights parity) and a new DayNightCompositor in rusty_sat_composites (band-major RGB blend, mask OR propagation, owned/borrowed).
• The integration test now renders three outputs: standard true_color (hybrid green, CIRA), JMA true_color_reproduction_corr (reproduced green), and the full JMA true_color_reproduction day/night blend (lim_low 73°, lim_high 85°) that replaces the bright gray limb with the natural uncorrected colors.
• White-background fix for the uncorrected composite: masked outside-scan pixels become NaN (the reader stores bright calibrated values there under the validity mask).

2. 2547b1a — JMA TCR enhancement (color conversion matrix + log stretch)

Root cause (land too green vs the JMA product): the JMA renders used the standard CIRA stretch. Satpy's true_color_reproduction_color_stretch chain is a per-pixel color conversion matrix (Satpy enhancements/ahi.py, Himawari-8/9) followed by a log stretch (min 3 / max 150, trollimage stretch_logarithmic). Added the fused finalize_rgb_jma_u8 finalizer plus JMA_CCM_HIMAWARI_8/9, jma_ccm_for_platform, jma_true_color_reproduction_value and log_stretch_value (trollimage parity), with hand-computed reference tests. Measured effect on Australian land: R−G +8.7 → +25.0 (warm red-brown, matching the JMA TCR look).

3. 6a497f5 — performance (byte-identical output)

Per-phase timing instrumentation showed the runtime split: corrections ~64 s (36%), a separate full-disk day/night-weight pass ~29 s, and two scene loads ~39 s. Optimizations, all verified byte-identical on the three output PNGs:

change	before	after
apply_corrections_with_sun_zenith_batch — five 1 km corrections share one per-pixel angle pass (bit-identical to sequential)	33.9 s	9.2 s
apply_correction_with_sun_zenith_and_weights — day/night weights emitted during the 0.5 km correction	24–29 s	0
uncorrected composite reuses raw-band clones instead of a second Scene load	21.9 s	0
trilinear interval lookup → binary search; secant clip angles hoisted per correction	—	—

Total: ~178 s → ~84 s standalone (2.1×), 305 s → 119 s under full-suite concurrency. The three output PNGs are byte-identical to the pre-optimization runs (verified with cmp), and the batch API has a bit-for-bit parity test against the sequential path.

Verification

• 679/679 workspace tests pass (release); cargo fmt --all -- --check and cargo clippy --workspace --all-features --all-targets -- -D warnings clean.
• New tests: Satpy angle-domain SZA reference values, out-of-LUT combined sunz-only parity, trilinear bit-parity vs rustyspectral, batch-vs-sequential bit-parity, JMA CCM/log-stretch hand-computed references, day/night blend semantics (day side == corrected, night side == uncorrected, gray-limb reduction, deep-night rescue).