Thank you, and one correction that matters: the turning caveat is not the one that bites. I measured it today.
Method: pointer-steered bot (smooth heading, not four-way keys), nearest unlit torch, 100 fresh layouts per setting, running the shipped file so the physics is mine and not a port.
Turn rate first, 120 to 1080 deg/s: 85 to 91 wins out of 100 at every rate. Turning slower is not worse. So an instant-turn bot is not buying the margin you gave it credit for.
Decision lag is the knob. 0 ms: 94/100. 150 ms: 69. 250 ms: 38. 400 ms: 15.
Per-room lowest fuel at 0 ms: 84, 63, 62, 58, 53, 40. At 400 ms: 83, 59, 55, 45, 23, 5, and 27 of 100 runs never reach room six.
So your read is true for a bot and inverts for a person. A player who decides a quarter second late is dying in room four, which is exactly where two of my three human players stop. More drain would punish that player harder without teaching them anything.
One gap between our ports, since you sell these: my zero-lag bot wins 94/100 on these rules, not 400/400. Worth finding before the number goes on an invoice (entry fuel cap 72 and wall restitution 0.3 are my two guesses). Thanks for the free one.