Opinion needed on ground planes for minimal boards?

I built something similar last year, a small module around a different micro but same idea, castellated edges and only power and IO broken out. What worked for me was accepting that L1 being mostly empty is fine as long as L2 is solid and unbroken under everything. The signals on L4 will use the plane right next to them, so your job is mostly making sure L2 has no slots or long splits where traces cross over. I dropped a handful of extra vias stitching the planes together around the edges of the board and near the ground pins, maybe a dozen total, and it made a visible difference on my scope. On the 8mil power traces, I did the same thing for the 3.3V rail and it held up fine, but I widened just the main feed from the regulator to a much fatter trace and kept the skinny runs short. The reference design guides are written for bigger boards, so take them as a starting point rather than a checklist.

One thing I would do differently if I started over: keep any fast signals, like the flash lines, on the shortest possable path and give them a clear lane over solid plane. The slow IO stuff is forgiving, the clock and flash lines are where I saw ringing before I tidied them up.

I'd push back a bit on leaving the top layer mostly empty and trusting the layer below to do all the work. On a module I made a couple years back I did exactly that, and had a stubborn analog input that picked up noise whenever the buck converter on the same board was switching. What fixed it was pouring copper on the top layer and tying it to the plane with vias wherever I could fit them. The noise on that line dropped enough that I stopped chasing it. The other thing I learned the hard way: any trace that crosses over a gap in the plane underneath will bite you. On mine, a couple of signals slipped over a via field that split the plane without me noticing, and the scope showed ringing on those lines only. With 8mil power traces you're already squeezed, so I'd pour what little top copper you have rather than leave it open. It costs you nothing in routing room since it fills around your traces, and it gives the high speed signals on the bottom something local to return through instead of making them reach up two layers.

One thing that caught me out on a similar tiny board was thinking my ground layer was solid when it wasn't. Vias crammed around the part had punched so many holes through L2 that the return path got pinched in a couple of spots. Flipping through the 3D view in KiCad and eyeballing the plane from below showed me the gaps straight away, way faster than squinting at 2D layers. Worth a look before you commit to the stackup, expecially with a part that dense in the middle of the board.

The point about vias pinching the plane is spot on. Same thing bit me on a small board I did a while back, looked fine in 2D but the 3D view showed a narrow neck of copper right under where I'd clustered a bunch of programming header vias. Shifting two of them over gave the return path way more breathing room. I'd also suggest checking the plane with the design rule checker set to report isolated islands, mine caught a tiny orphaned patch of copper I'd never have spotted by eye.