Here is a repro for String#[]=:

length = 10*1024*1024
buf = " " * length
length.times { |i|
  buf[i] = "0"
}
p buf[0]

and the backtrace:

$ jth ruby --graal 16bug2.rb                                           
.../jruby-dev/truffle/src/main/ruby/core/string_mirror.rb:65:in `splice': stack level too deep (SystemStackError)
	from .../jruby-dev/truffle/src/main/ruby/core/string.rb:1308:in `[]='
	from 16bug2.rb:4:in `block in <main>'
	from 16bug2.rb:3:in `times'
	from 16bug2.rb:3:in `<main>'

I'll mention this old pull request since it might have been trying to solve a similar issue:
#4201

This is unfortunately a pathological case for ropes currently. Replacing each byte or character is problematic because the ropes themselves are read-only. Preallocating really has no additional benefit because you're not going to update bytes in a fixed buffer. But, if you really just wanted 35 MB of "0", then String#* is your best bet.

I'll see if I can find a way to make this work better.

At first I do not care so much about performance, but at least it should not stack overflow.
My guess would be it should be a loop for looking up a character in a rope and not recursive calls.

The first snippet only does appends, so that one should not be so bad with ropes, isn't it?

It looks like 8abc05a solves both cases, thank you!
The full script, using << is reasonably fast but slower than MRI, and surprisingly the version using []= is about the same speed, while I was not patient enough to wait for MRI to terminate.