For me it simply looks like the reason is that sampling rate is below minimum advised value of 5 MHz (TX filter-related limitation). Check this topic. People have reported that using LO offset could help a little, but personally I would try to avoid all these decimations etc. and just check higher sampling rate in traditional approach. If you will specify the bitrate (I am not sure if I see it right that you want it to be 32 kb/s, yielding double-sided BPSK signal main lobe width of 64 kHz), I could try to achieve the same effect with my way of generating of such kind of signal.
//edit: this is what I got after playing with GRC for a moment. It clearly shows that these distortions may be avoided.
Sorry, but this is not true at all. In theory, board is supporting sample rate up to 61.44 MS/s and though it will probably start to drop samples for that high sample rate, there is no reason why you cant successfully use sampling rate of, say, 45+ MS/s on TX side.