Excellent testing technique.

But Im sorry.... my bad. I only skimmed over your get-heights,
so I didnt see the ent_getvertex's in there.
They are a well-known SERIOUS issue with speed when calling on a pix-by-pixel basis.

As you are certainly aware, when you just call ent_getvertex(?,?,xx) you get
just ONE vertex worth of data. Obviously..
But were you aware, if you call ent_getvertex(?,?,1) then ent_setvertex(?,?,1),
it gives you access to ALL the vertices, without any more ent_getvertex calls?
(At least until the next wait(), and then the buffer gets flushed away.)

I know your code, to my un-familiar eye, doesnt look friendly to the idea,
but is it PLAUSABLE to fit the following concept into a SINGLE frame...?
[because if you hit any wait(), you need to re-call ent_getvertex(?,?,1) & ent_setvertex(?,?,1)]

1> Call ent_getvertex(?,?,1) then ent_setvertex(?,?,1) to access the vertex buffer.

2> Set all your heights DIRECTLY into this buffer. Do a ent_setvertex to lock it in.
[note: even though you did a ent_setvertex(?,?,1) to save, the buffer is STILL THERE until the next wait()]

3> Paint your skin using the data from the buffer rather than individual ent_getvertex calls.

Does that sound within the realms of possibility?
And YES, I am aware thats a LOT of changes...

I am CERTAIN this will give you a serious improvement, but I GUESSTIMATE
it will actually give a HUGE improvement, based on my rough understanding
of the sheer number of times your get_height gets called, which looks large.
BUT, will it be enough? And will the improvement be worth that much re-coding?

I certainly HOPE so.... But only time and MUCH effort will tell...

Best of luck...