Maybe I'm wrong, but as far as I remember Blitz3D doesn't support Render to Texture. The addition of Flag 256 allowed for faster access ("Store texture in vram"), but no direct render to texture. Copyrect from backbuffer to vram-texture is rather fast tho. Lockbuffer is slow and the actual problem. A way around would be:
render actual scene, without rays.
render mini version 256x256 and copyrect to texturebuffer (optimized with details like grass hidden)
render rays mesh alone ontop
This 3 render method does not allow for pixel access, so I'd use EntityColor 0,0,0 for shadow casting things, like in the source. Most important: no lockbuffer required. That should lower the 16 ms to 2 ms. The beauty of extracting the texture from the first render using readpixel etc. was, that its speed is independent from scene complexity, but well, it required lockbuffer, so..
edit: I just tried that, not significantly faster, unless I did something wrong. Then there was also a hack, to directly peek/poke VRam and Buffers, I guess using the memory.lib. Most likely not very stabile.