Reproduced in:
Version 2020.1.0a2 (d3fe9dc000c3), Sat, 24 Aug 2019 19:47:28 GMT, Branch: trunk
Version 2019.1.12f1 (f04f5427219e), Wed, 17 Jul 2019 18:27:29 GMT, Branch: 2019.1/staging* (results were different, but worse. see buttom)

Summary:
When using a multi-tile terrain system with large heightmap data (note: not splatmap!), when the user makes a large texture painting stroke (altering the splatmap), undo can silently fail for most of the affected terrain tiles (but not all).
Presumably a unfortunate consequence of that fact that the splatmap and heightmap data is all in the same file.

Repro:
1) In a new scene, create a terrain.
2) In the terrain settings, change the Heightmap resolution 2049 x 2049.
3) Using the Create Neighbor tool, make 5x5 Tiles starting with this first terrain tile.
4) Add a texture to all of the tiles, and then create a 2nd terrain layer.
5) Now, as seen in the .gif, snake up and down the 5 tiles using the 2nd texture.
6) Undo.
> It will only undo some of the changes. No error.

(The rest is optional; only shows that this is a problem for high res terrain).
7) In a new scene, create a new terrain. Don't change the heightmap resolution from 513x513.
8) Repeat the same steps.
> Will undo.

- Also reproduces for Heightmap data.

Expected:
- At the least, silent undo failures are undesirable. It's better that the user is informed that what they are doing cannot be undone. Preferably before they do it, but something during the undo event would suffice.
- That the undo fails completely or not at all. Undo only some of the terrain tiles just creates more problems.
- From the user perspective, it's unexpected that this would be the case since the splatmap resolution is the same for both of these terrains. Why is the heightmap data affecting my operations when changing the splatmap data? Ect.
- That the undo for splatmap is recorded separately from the heightmap. But they are using the same file after all...

* in 2019.1, it was doing the same except none of the terrains would undo the changes, as opposed to just some. Perhaps because of Chris's changes involving undo large terrain data. So the bug has changed over time, but not a regression.