Either document them or remove them from the language.
If they are removed from the language then the code in tokenizer.cpp that doesn't call bigfloat_init_buf can be deleted. If they stay in the language, zig should probably reformat the string to match hex float literals and still have the same float parsing code handle them.
i vote for removal
The grammar changes are done thanks to @mokulus.
As to the stage2 parser, I think the tokenizer can allow octal and binary float literals, and then semantic analysis can reject them as invalid format. We already defer float parsing to semantic analysis in stage2. So this can be considered closed.
Most helpful comment
i vote for removal