The BeanShell version of that last test ran at 8 fps (where the pure Java version runs at 12 fps and the JRuby version ran at 3 fps). So the slowdown isn't too bad, plus the clincher is that I can copy-paste any performance-critical BeanShell code into java.
(ps, the reason those FPS numbers are so low is that the test is re-rendering the SVG tree on every draw. In the real game, if I do use SVG sprites, they'll be pre-drawn and cached)
Plus BeanShell gets a heap of bonus points because it worked the first time, and behaved just like I expected. I can't say the same for those other engines :)