Every time read_blocks_by_block yields, the pixels array has the size of block_buffer_size.
Given previous code:

          pixels.each_slice(x_block_size).with_index do |row, block_row_number|
             yield row
             break if block_row_number == y_block_size - 1
           end

If it is the last block, the x_block_size is actually the remainder value.
This causes the pixels to be sliced incorrectly, iterating more times than necessary and reaching the break condition faster, discarding valid block pixels.

Would you mind adding a test that verifies the fix?

Sure! Should be ready in a couple days

@turboladen here it is.
I had to also fix write_xy_narray, otherwise I'd have to workaround it by writing into the raster in a different way.

It kinda had the same mistake, when writing the last block it would ignore the remaining block pixels and cause the final raster to have missing pixels.

Note: Both GTiff driver and TILED options were used on the new test to ensure that the created raster would have a block size greater than the raster size itself.

Added the suggested changes. New code now uses Numo::NArray instead of NArray

All specs are passing locally 😄