Unfortunately, the video link above doesn't work for me, so I can't see the "bad effect" that you mention. It says the video is unavailable because it is private.
Although the images are nice, I'm not exactly sure what your desired goal is. I can guess that 1 is bad because it makes some elements too small and that 2 is bad because it abruptly cuts the tiles at the end. However, I don't understand what is wrong with 3.
Questions:
(1) What is your desired output for the red, green, and blue tiles when your facade width is short, medium, and long?
(2) What is the cga code used to generate your examples 1, 2, and 3?
If I were to make a guess, I think you have a pattern which should be about 15m, and you would like to repeat this pattern across a facade of any width such that the pattern stays as close to 15m as possible. The ratios between the tiles in the pattern should also be kept. The pattern consists of a blue tile on the left, a green tile in the middle, and a red tile on the end, where the green is about 75% of the size of the red one, and the blue is about 50% of the red one. In this case, I would use a repeating split that splits out pieces that are approximately 15m (your pattern width). Then, for each pattern piece, do another split that splits it into blue, green, and red tiles.
Facade -->
split(x) { ~15: Pattern }*
Pattern -->
split(x) { ~0.5: color(0,0,1) Blue
| ~0.75: color(0,1,0) Green
| ~1: color(1,0,0) Red }
In order to avoid extremely small tile pieces, you could also put a case statement in the Facade rule to check if the width is larger than some minimum threshold before doing the split.
