Issue reading LZW TIFF with HORIZONTAL_DIFFERENCING predictor.

[palisades-lakes]

I am having problems reading a NASA tiff image.

A minimal example is in https://github.com/palisades-lakes/imageio-ext-issues

See https://github.com/palisades-lakes/imageio-ext-issues/blob/main/src/main/java/iiox/Lroc.java.

It does a read-write-read roundtrip equality test.
With or without the imageio-ext-tiff dependency, roundtrip equality succeeds.
However, visual inspection of output image:
https://github.com/palisades-lakes/imageio-ext-issues/blob/main/images/lroc/lroc_color_poles_2k-iiox.tif
shows it is corrupted.
Output produced with the imageio-ext-tiff dependency removed:
https://github.com/palisades-lakes/imageio-ext-issues/blob/main/images/lroc/lroc_color_poles_2k-iio.tif
looks correct.

[palisades-lakes]

Looks like the problem is PredictorDecompressor.decode() line 64:

existing version:
int count = bufOffset + samplesPerPixel * (j * srcWidth + 1);

seems it should be:
int count = dstOffset + samplesPerPixel * (j * srcWidth + 1);

This fixes the roundtrip for lroc_color_poles_2k-iiox.tif.

Existing unit tests all pass either way, which suggests something is wrong or missing wrt testing LZW horizontal differencing.
A very superficial look at the test images leaves me wondering if some are all zeros...