Thanks for pointing this out!
The big change I made is that I now write out the alpha channel, which requires using
the BITMAPV4HEADER. The documentation is correct: the compression method must be set to BI_BITFIELDS in order to specify the alpha channel otherwise if 32 bits-per-pixel are specified with a compression value of 0 (BI_RGB), alpha is ignored. I only tested with IrfanView and everything looked fine but you are correct that things look wrong in MS Paint as well as Gimp. Hmm.
Well, what's happening is that some textures have the contour bit (the T bit) enabled in various pixels but when the textures are used in the game, contour processing is disabled on those polygons. The way the T-bit works is that when set, the pixel is transparent (so I write alpha = 0 when T = 1), but this capability can be controlled on a per-polygon basis by the mesh data. So even though there is one T1RGB5 format, there are in fact two ways of interpreting it. I've fixed the problem by writing out two T1RGB5 files.
If you open up the file in Gimp and look carefully, you will notice that the AM2 logo background in Daytona 2 is cut out (transparent) as it should be with contour processing enabled. But by changing the compression method back to 0, this is no longer the case. Also, if you open up e.g. the RGBA4 textures, you'll see the alpha channel is gone entirely.
Revision 879 should be built sometime tonight.
Glad to see you back in the commit logs Bart. It's been a long (almost) 4 years. It's not the same without you.

It's nice to be back. I wish I had more time to contribute

I was really irritated that Apple never granted me permission to work on Supermodel. Reviewing open source project requests was not a priority for the company. I got a call more than a year and a half after I submitted my request from an apologetic lawyer claiming they were moving through their backlog of hundreds of OSS requests but then I quit 6 months later.