Upon closer inspection the error signal seems to vary quite significantly on the scope (scanning @ 2 Hz), sometimes completely flipping its sign even though it always triggers on the same side of the ramp (see attachment for video, along with some neck excersise).
This might be the same behaviour from before, whereby the demodulated signal might still be "riding" a low-freq componen which can't be compensated with the LO (Marconi's carrier resolution = 1 Hz). Using the 10 MHz external Rb reference doesn't change anything. It seems that even with the coupler, reflections may be entering the mixer...
Adding a LP filter (BLP-1.9+) right at the mixer output solves this for good. Even using 36 MHz LO vs anything else doesn't make a difference so this explains the previous issue. Moving back to lock using stable err signal.
For reference, the LO carrier is set to 36.000 MHz, +7 dBm (so the EOM is driven with an estimated +30 dBm well below the saturation or damage threshold +40 dBm).
Achieved a good lock for pretty much all of the afternoon today. The laser ran at 937 mA current, the optimal gain on SR560 was found to be 50, with a LP cutoff at 300 Hz (12 dB/oct rolldown). The 300 Hz cutoff supresses most of the nasty 8 kHz noise (and harmonics) which I can hear with enough gain. Source still to be determined.