Well it is loosing a lot of height, it is mixing to post-eruption. It is the same mountain and resolution LIDARs at same position, so not much will really happen in a 41 year span.
When doing something like thousands of years you're going to need multiple merges and phases.
Merge one mixes to A 100%, merge2 then mixes to A 100%, then merge3 mixes to A 100%, etc, setting the keys accordingly. Each erosion map you mix to should have gradual changes. More down cutting for example, with progessively more duration (though maybe only stepping by 0.05 each time).
Using the erosion shader itself seems like it maybe problematic inherently because of it being raster based. But maybe if you did less frames, and interpolated between them in post you'd get less dancing.
Maybe the plugin also responds to lowering the CPU count within TG? Maybe 1 CPU would provide a more stable map. But naturally be much slower.
PS in my example, you can see though, even with the features being mainly the same on the sides the mountain didn't slip, the tinest change in shape causes flows to totally flip sides of displacement, taking on whole new courses. So if the terrain changes any over time, this will dramatically impact the erosion result.
