Leakage from interference of seemingly equal intensities

I attempted to design a 25/75 beam splitter and ran into a strange interference problem. I split the initial photon into 25%, 25%, and 50% beams without issue. I also managed to rearrange a 25% and 50% beam into two 37.5% beams (of equal phase). However, no matter how I try to merge those two beams through interference, I always end up with about 73% from one side of the half-silvered mirror and 2% from the other side.

At first I thought that one of the components was absorbing a fraction of the beam, causing the intensities to be slightly off. To that end, I've tried using a Mirror and a Coated 50/50 Beam Splitter; a Glass Slab, a Mirror and a 50/50 Beam Splitter (in case the Coated 50/50 Beam Splitter is to blame); and even some Quarter Wave Plates, Double Sugar Solution, Mirrors, and a 50/50 Beam Splitter (using circular polarization and sugar in case the Glass Slab is to blame). Each time, the result is the same: 73% leaves in the expected direction while 2% sneaks off at a right angle.

It seems like either there's something fundamental about interference that I'm missing, or there's a bug somewhere.

Analysis: The game reports the two 37.5% beams as (0.500 + 0.354i) and (-0.354 - 0.500i), giving them a intensity of 0.375316±0.0008545. It reports the 73% and 2% beams as (-0.604 - 0.604i) and (0.104 - 0.104i), giving them respective intensities of 0.729632±0.0012085 and 0.021632±0.0002085.

Below is a transcript of the quantum state as reported by the status message. Here is a copy of the level (pasted as .txt because Github doesn't like .json uploads). Any ideas?

(1.000)*|0,9,>|)
(0.707)*|1,9,>|) + (0.707i)*|1,9,^|)
(0.500)*|2,9,>|) + (0.500i)*|2,9,^|) + (0.500i)*|1,8,^|) + (-0.500)*|1,8,>|)
(0.354)*|3,9,>|) + (0.354i)*|3,9,^|) + (-0.707)*|2,8,>|) + (0.354i)*|1,7,^|) + (-0.354)*|1,7,>|)
(-0.354)*|4,9,^|) + (0.354i)*|3,8,^|) + (-0.707)*|3,8,>|) + (-0.354i)*|1,6,>|) + (-0.354)*|2,7,>|)
(-0.354)*|4,8,^|) + (-0.500)*|3,7,>|) + (-0.707)*|4,8,>|) + (-0.354i)*|2,6,>|)
(-0.354)*|4,7,^|) + (-0.500)*|4,7,>|) + (-0.707)*|5,8,>|) + (-0.354i)*|3,6,>|)
(-0.500i)*|4,6,>|) + (-0.500)*|5,7,>|) + (-0.500)*|6,8,>|) + (-0.500i)*|6,8,^|)
(0.500i)*|5,6,^|) + (-0.500)*|6,7,>|) + (-0.500)*|7,8,>|) + (-0.500i)*|6,7,^|)
(0.500i)*|5,5,^|) + (-0.354)*|7,7,>|) + (-0.354i)*|7,7,^|) + (-0.500)*|8,8,>|) + (-0.500i)*|6,6,^|)
(0.500i)*|5,4,^|) + (-0.354)*|8,7,>|) + (-0.354i)*|7,6,^|) + (0.500)*|9,8,^|) + (0.500i)*|6,5,>|)
(0.500i)*|5,3,^|) + (-0.250 + 0.354i)*|9,7,>|) + (0.354 - 0.250i)*|9,7,^|) + (-0.354 - 0.250i)*|7,5,^|) + (0.250 + 0.354i)*|7,5,>|)
(0.250 - 0.354i)*|10,7,^|) + (0.354 - 0.250i)*|9,6,^|) + (0.354 + 0.250i)*|7,4,>|) + (0.250 + 0.354i)*|8,5,>|)
(0.250 - 0.354i)*|10,6,^|) + (0.354 + 0.500i)*|9,5,>|) + (0.354 + 0.250i)*|8,4,>|)
(0.250 - 0.354i)*|10,5,^|) + (0.354 + 0.500i)*|10,5,>|) + (0.354 + 0.250i)*|9,4,>|)
(0.500 + 0.354i)*|10,4,>|) + (-0.354 - 0.500i)*|11,5,^|)
(0.104 - 0.104i)*|11,4,>|) + (-0.604 - 0.604i)*|11,4,^|)
(0.104 - 0.104i)*|12,4,>|) + (-0.604 - 0.604i)*|11,3,^|)
No goals, no judgement.

stared / quantum-game

Leakage from interference of seemingly equal intensities #18