He either made his judgement based on non-obvious criteria or, more likely, looked at x265 at the "wrong" moment. Right before libbpg was announced, x265 had several commits fixing large 10bit and 422/444 bugs.
As far as intra-prediction goes, x265 should be better than the HM by at least a few percents MSE-wise and around 10% SSIM-wise[0].
[0] https://github.com/strukturag/libde265/wiki/Intra-Prediction...
I'm interested in seeing the result of using 4:4:4 with BPG. If you have the originals to compress, of course (meaning, the original is not chroma subsampled).