Interesting measurements!
On my own build VNA using exactly the same principles as the NanoVNA (SI5351 and 3*SA612) but using a PC audio input instead of an on board adc and a PC processor for the dsp work instead of the STM I observed the audio output of the transmission SA612 using a realtime 16k points FFT. This makes the noise versus the 5kHz signal very visible, it clearly shows the frequency dependent leakage where the real noise is independent from the LO signal to the SA612.
The pictures I get are similar to what you get but due to the much longer FFT and thus much smaller bucket size the noise is at -110dB and the leakage signal varies between below noise (sub -110dB) below 100MHz till -50dB when around 900MHz. A full thru signal is at -10dB.
Of course a dead bug style bridge will have lower performance compared to the well done PCB layout of the nanoVNA but the leakage amounts are very similar
