The algorithm is rather noncursed: always send the atom on the counterclockwise side of the missing bond first and go counterclockwise. With this, the bonding station does the same thing in all cases and result in the same bond shape when it fulfills the leaving condition (well, to save regrabbing instructions it leaves bonder as full triangle but the desired bond shape is the same relative to the bonder glyph so it fixes on the same debonder on all cases), and the difference between cases is just rotation, and can be arranged so that all three cases finishes rotation at the same time.

Now, to achieve this conduit signature, the trick is to artificially add bonds to one case so that the three cases enter the debonding station 2, 3 and 4 times respectively, and thus the machinery has to deal with 4 cases. The 1-atom-left case and the 4-atom-left case both has a lone atom on the far end of the debonder and the only way to differentiate is whether there's a 2-stick on conduit, so the lone atom has to rebond to the 2-stick. The 3-atom-left case and 4-atom-left case both have three bonded atoms going back from conduit to debonding station, but the 4-atom-left case must redo lone-conduit bond while 3-atom-left case must not, so two arms (arm 3 and arm 4) are needed to execute the conduit-to-debonding-station move in two separate paths.

The design is not entirely trackless but gives out a lot of TI vibes like 1/6P and rotate-back-and-forth. The instruction count in debonding chamber, 17, isn't exactly low as I imagine there can be some one-big-track or one-big-track-with-helper motifs but don't have the expertise to go deep into. I also have a very cursed debonding chamber that does a debonding job in 16 and 1/2P but its conduit signature probably takes 30+ instructions in the bonding chamber. Still, looking forward to a (debonding-chamber-instruction bonding-chamber-instruction) pareto frontier mapped out.