Thanks for the prompt reply Miti!
Indeed, my 'analog wire' apparently was too long (and therefore noisy?).
To make connections as short as possible I selected A7 as sample port,
soldered the components on the Arduino board (see picture below, I use a BAT86 diode)
and routed an extra GND wire from Arduino GND to HC4046 GND (pin8).
The result is flabbergastingly different in a positive sense!
Until this moment of writing it never missed a beat.
Edit some hours later: it still hasn't
In run mode the DAC value converged and clamped to 0 (zero), so I had to
to lower the minimal voltage at the VFC pin and used 100k parallel too.
Edit: changed the resistor to 51k1 (so that DAC value appears ~31000 when in lock)
Now with h1 and h65535 the difference in diff_ns was around 600, so my gain is 65536/600 = 110.
Edit:with 51k1 resistor h1 = 996 mV, h65535 = 2200 mV --> Δ diff_ns = 490 --> gain = 134, I took 128.
Letting it run, the contraption locked within 10 minutes
Edit: in warm condition (with 51k1 parallel to VFC) within 130 seconds.