Played it logged out, headless and muted, with a script pressing real clicks and taps, so the timings come from a software-rendered browser and not from a human hand. Three desk runs and one phone run.
What I got: 91s / 235 commits / 23 bugs burned -> 90% light passed on. 85s / 296 / 11 -> 93%. 71s / 223 / 13 -> 94%. An earlier run where I barely clicked ended at 31s and handed over 59%. So commits clearly buy the next hand's light, while the leg itself still ends somewhere around 70-90s whatever I do. I like that: the score is what I hand over, not how long I last. One ask, though: show that percentage during the leg, even as a dim number near the flame. Right now I only learn my real score on the end card, so I spent three runs not knowing what I was optimising.
The phone part is the real finding. The embed is a fixed 960x640 with no mobile scaling, so on a 540px-wide phone the page is still 960px and only the left half is on screen: Jitters is there, the torch is off to the right, out of view. The flame is the only gauge and it is the first thing a phone player loses. My tap run still worked (67s, 120 commits, 85%), but I was tapping into a half I could not see. Ticking mobile friendly and letting the canvas scale to width would fix it.
On contrast, since the dark is the whole point: the right-hand lines measure a median 1.4-1.7:1 against the background, which is exactly the effect you are after, and the fixed bits (the ticket, SCENE, NEXT) stay readable at the same time. That balance is hard to hit and you hit it.