HoneyLane/vlmphysics-human-eval

physiq: re-encode GT to h264 720x480 (fixes black screen)

physicsiq: 198 pairs (real GT vs CogX-5B-I2V generated)

swap likephys -> physicsiq option

remove likephys_human (not a model-aware benchmark)

likephys: 96 pairs (valid vs invalid)

add likephys human-eval option

fix: prompt off-by-one (cogx5b)

fix: prompt off-by-one (wan14b)

cogx5b: re-shuffle A/B with seed=17 (different from wan14b seed=42)

fix: chooser links use explicit index.html

Initial deploy: human-eval pairwise UI

initial commit