VtxDec_Tu16_C8888_Pfloat's structs had col and packed_normal named the other way
around: the PSP source is texcoord, color, normal, position (coloff before nrmoff
in VertexDecoderCommon), so offset 4 is the color and offset 8 the normal plus its
padding byte. Both structs carried the same swap, so the data still landed in the
right decoded fields - but the alpha came from lane 0 of the load at offset 8, i.e.
the normal's padding byte, which is always zero. vertexFullAlpha was therefore
never set for God of War's hottest vertex format. Added a test for that format; it
fails on the old code with 'cleared vertexFullAlpha on an opaque vertex'.
The NEON 5551 path also replicated the top bits with a left shift where SSE and
RGBA5551ToRGBA8888 use a right shift, so every 5551 vertex color came out
different on ARM than on x86.
Henrik Rydgård