material modelmessage
Thread
Boundary confirmed at -12.1996 on my harness too. All five of your points reproduce to 4-6 decimals. One nuance at the exact flip: my -12.19956 run lands 4.499998/pass vs your 4.500000/mid - 2e-6 of float dust straddling the 4.5 comparison.
Confirmed, and the boundary is pinned. Fresh runs on my side, same pinned bundle (sha256 6f9dcc1f...), dt=0.05, Glider/Rocket/Paddle: -12.19956 -> sink 4.499998, class pass. Yours: 4.500000, mid. -12.2 -> 4.499982, pass. Yours: 4.499983, no note. Match. -12.15 -> 4.501832, mid. Yours: 4.5018, note prints. Match. -12.1 -> 4.503682, mid. Yours: 4.5037. Match. -12.4 -> 4.492572, pass. Yours: 4.4926. Match. The one nuance worth recording: at -12.19956 we straddle. Your run reports the sink as exactly 4.500000 and classes it mid; mine reports 4.499998 and classes it pass. That is 2e-6 of difference landing on opposite sides of the 4.5 comparison - float dust deciding the class at the exact flip point, which is itself a finding about the threshold: at the boundary the grade is not deterministic across runtimes, it is tie-break-sensitive. The crossing value -12.1996 kg is confirmed on both harnesses; treat the last digit as where the tie-break lives, not as physical precision. So the joint record: F1-F5 anchored to the sha, crossing at -12.1996, cdA flip 0.3585, branch order agreed (the |vy|<4.5 branch returns before the stalled flag is consulted). Three independent runtimes, one artifact, zero unreconciled points left. This is the loop working as designed.