I wouldn't worry about feedback path length, at these frequencies it could be meters before you'll notice. Assuming noise or other feedback doesn't couple in along the way, of course (hence why planes are so great), and assuming the path length isn't on the feedback node itself (which shouldn't be loaded by much capacitance, but this also can be compensated with more "speed-up" cap).
I was actually proud that I found a clever way to keep the feedback loop short by flipping the 3886
... I was completely unaware that this is a bad practice.
Not sure what will happen if I remove the flip - but the feedback loop will definitely be longer. Are there any reasons except aesthetics and volume?
The heart of the argument I think is more about manufacturing and space. If those are unimportant (as a low quantity hand-built is), it's meaningless.

I feel a kneejerk about it because it's very dissonant against what you see out there, a style issue -- this is however the weakest possible argument (and being honest enough to notice when this is the case, is a challenge unto oneself).
But as for useful justification, it's a design-for-manufacture (DFM) issue.
In production:
- You'd have to hand-solder at least one pair of amps. No chance of running the board over a wave solder machine twice.
- Heatsink screws on both sides is fiddly, and makes order of assembly that much more critical. (Probably can't mount the board before finishing those screws?)
- Requiring access means whatever mounting for the board (standoffs/bosses) has to be very tall.
On the other hand, all the same... if you're going for an industrial chic look, it might be a feature-not-a-bug. Say you tilt the whole board so the LM3886s look like they're in rows or something. Idunno.
In the latter case, fiddliness of assembly is probably a feature-not-a-bug, too.

Yes, the low-power ground was missing a connection to the high power one. I intended to make this connection on the PSU-PCB but since I don't think there are big currents on the ground traces (as it's a bridged amp and the speaker doesn't return to ground) I think I will just connect them on the main amp PCB.
Yikes! Don't run a critical signal through a wiring harness!
Incidentally, be very weary of advice you see about star grounding. There are some very beautiful (in the industrial-chic sense) implementations of this out there... and they stink to high EMI heaven! Fortunately, audio amplifiers are extremely lenient, so that most any beginner can put together a schematic (with pretty much any layout) and get acceptable results, at least assuming they aren't in an RF warzone.
The better way is simply to route current paths, and signal loops, away from each other. With high frequency currents returning on the ground plane underneath the trace, loops are where you put them. In this case, it should be easy to locate the speaker terminals (if this were an unbalanced amp) near the power amp, and have the signal amp on the other side of that connection.
Also, it's just clicking that this is intended for bridged operation... so you could have some opportunity to measure the
differential voltage, and feedback with that. A differential sense amp, that's easy, but maintaining balance is actually a bit harder.
A fully differential amp could deliver +/- output (feedback) signals, but, I doubt one exists that has nearly as good specs as the '49710.
You could feedback to one phase (the "slave" or "follower"?), and leave the other (the "master") alone (similar to what they did back in the day: the paraphase phase splitter). That is probably a bit better than the split path method, but it feels inelegant.
Point is, you're approximately doubling the errors in the system -- the errors from both outputs add (well, subtract, but assuming random variables y'know). This is not wrong to begin with -- you've doubled the output power too, after all!
Using a single error amp, would reduce it to a single error, not a double error, so it's worth about 3dB SNR. (And, again, not that this is really needed, but given the spec of the '49710, I assume you're going for the best possible.)
So, it's merely an incremental improvement, versus, there may not be any good way to implement it with off-the-shelf parts.
I don't think I would do anything different. This is more to illustrate design analysis, the kind of contemplation and insight that any good design needs!
I played around with ground planes but I wasn't sure how to go around it. From what I've read, the best way seems to be to use a ground plane for the signal ground and route the power ground by hand.
The new PCB does this (still WIP) on both the top and bottom layer. I still have to do some serious via stitching.
Fixed another issue while was at it: the output zobel inductors were parallel and they could couple magnetically. Now they are 90 degrees apart.
Ah, cool.
FWIW, nearby (air core) inductors typically have k < 0.1, which isn't much, and I doubt you'd notice anything here.
Actually, the ones on opposite phases of a given channel, are getting the same current anyway, so they act in series and the coupling is equivalent to a slight increase or decrease in total inductance. Should be no problem there. The total value would be slightly different from what's intended, but these are low-Q parts, errors over a factor of 2 would be cause for concern.

Crosstalk between channels, via inductor, isn't going to be noticeable until near cutoff, either; keeping them isolated, really just feels better, and that's fine.
Good point about the holes. I will make them bigger and remove the ones on the amp side.
The LM3886 footprint seems to match the datasheet spacing between the pins and the heatsink - so no extra space there. For now I don't need that extra routing space but good tip with the heat spreader. I'll think about routing power through there, see where that leads. Maybe it solves some of the ground place issue.
Actually, y'know, leave the holes in, in case they're useful -- see how it goes in the real build. Don't HAVE to put standoffs and screws in them. If nothing else, maybe they'll prove useful for wire ties? Who knows.

That's another nice lesson -- prepare for options. Not many options to worry about on an amplifier, but it can be nice to have a number of footprints for testing different components and combinations when you expect you have a design that's going to need some fiddling.
Or... conversely... if you're not averse to defacing a PCB, there are many things that can be changed on an already-fabbed board, for varying degrees of difficulty. Cutting traces and running wires is easy. Drilling holes -- if you have free space to do so without cutting or shorting anything out -- is pretty easy, too. Patching traces in a like-new manner, is more difficult, but can be done. Rebuilding board material entirely, well... So, even if you don't put holes in the design, they can be added (cough... subtracted?) later, if the space is still reserved for them.
Tim