It is fixed in libXft 2.3.6:

https://gitlab.freedesktop.org/xorg/lib/libxft/-/blob/libXft-2.3.5/NEWS

* adds missing function prototype
* move xgetcolor() prototype to win.h (that's where all the other x.c
  func prototype seems to be declared at)
* check for snprintf error/truncation
* reduces code duplication for osc 10/11/12
* unify osc_color_response() and osc4_color_response() into a single function

the latter two was suggested by Quentin Rameau in his patch review on
the hackers list.

the array is not accessed outside of base64dec() so it makes sense to
limit it's scope to the related function. the static-storage duration of
the array is kept intact.

this also removes unnecessary explicit zeroing from the start and end of
the array. anything that wasn't explicitly zero-ed will now be
implicitly zero-ed instead.

the validity of the new array can be easily confirmed via running this
trivial loop:

	for (int i = 0; i < 255; ++i)
		assert(base64_digits[i] == base64_digits_old[i]);

lastly, as pointed out by Roberto, the array needs to have 256 elements
in order to able access it as any unsigned char as an index; the
previous array had 255.

however, this array will only be accessed at indexes which are
isprint() || '=' (see `base64dec_getc()`), so reducing the size of the
array to the highest printable ascii char (127 AFAIK) + 1 might also be
a valid strategy.

all the ctype.h functions' argument must be representable as an unsigned
char or as EOF, otherwise the behavior is undefined.

Build on auto-sync and only mark window dirty on palette changes and let
the event handler do the actual draw.