Those are all excellent points. However, I must ask: what part of my post was inaccurate, exactly? Could you quote it for me?
When you say "forcing your choice on users is IMO not wise", do you mean the FreeBSD devs? Because as far as I go, I stated that "having the very choice of picking one or the other" is what really matters, regarding endianness, meaning I'm entirely with you there.
Anyway, I don't think all those points, which seem to sum up to "better GPU support" and "more compiler assumptions", invalidate any of the previous points. The GPU situation, in fact, goes hand in hand with my earlier statement, "we still live in an x86 world", and is definitely a very good argument pro-little-endian.
But I must say that although it doesn't concern earlier PPC hardware, because they can't use modern GPUs due to being stuck to earlier revisions of PCIe (and many still-great big-endian GPUs are available for them), it certainly concerns Raptor computers with their amazing PCIe 4.0 interface, and that's a point I hadn't realized before (modern GPU endianness), so I thank you for bringing that up. Of course, nothing prevents all of it to work just as well with big-endian (flipping byte order aside, unless if new BE GPUs are made), but as you pointed out, new work would be required, and who would volunteer to do it? With this, the thread title finally sort of makes sense in at least one point (which otherwise doesn't, as either big or little-endian are equally Talos-focused).
Also, just to make sure I understand you correctly, when you say "and even if stuff gets implemented it will always be slower/less efficient", you exclusively mean the byte-order flipping from little to big-endian in modern GPUs, right?
Regarding compiler assumptions, on the other hand, I don't think that is much of a point, because you can always compile assuming CPUs of a specific "range" or "category" are being used regardless of endianess. For example, you can always compile and assume POWER8 and later is being used, with AltiVec, VSX and all the latest POWER ISA, and still choose big-endian, no runtime checks required. There's absolutely nothing stopping the developer there, as far as endianness is concerned. In fact, that's what Fedora ended up doing, around version 24 or so: they offered the system for big-endian PPC systems, but they weren't, say, fully G5-compatible anymore, even though it has AltiVec and so on; Only later CPUs were officially supported.