I tried building the circuit using the 4"/2mm vials from aliexpress. I used the regular adhesive-backed copper tape again; I think it's probably good enough. I made the two top electrodes a little narrower than the bubble. This time I used 1.8432 MHz 14-pin DIP can squarewave +5V oscillator from the junkbox plus a 74HC buffer to drive the circuit, and made the two 0.01 uF caps 0.002 uF instead, all the rest was the same. My crude screw-on-end-of-lever arm calibration indicates the full output range of -220 mV to +240 mV covers just 31 arc-seconds. In quiet moments I see about 80 uV of noise which would correspond to about 5 milli-arc-seconds. Not sure if my calibration correct but if so, that's plenty sensitive for my needs. This vial seems to take over 10 seconds to respond after a step change, much slower than the smaller and less sensitive, 1.5 inch long antique vial I used to start with.
EDIT 11-June: A quick test indicates the vial is roughly twice as sensitive- on average- as claimed. Added output graph. My above noise estimate almost certainly optimistic. After a 7-hour run overnight, I see the output has sudden jumps as large as 6 mV, or 0.4 arc-second per my rough calibration. The bubble will track infinitesimal level changes only if the interior is absolutely smooth. My guess is that the supplier did not waste time making the interior surface of this $12 vial any smoother than needed for a simple visual readout, so my circuit is showing the surface roughness as steps in the response. Also possible these were cheap because they were rejects. It's a little disappointing, but not entirely unexpected.
EDIT 2: The vial was not cleanly or securely mounted, so I cannot rule out the step changes overnight as related to thermal expansion + roughness of my own mounting surface. I will try to make a better mount later.