节点 n30

CHANGE: α from 1.5 to 1.75 (interpolating between the two best-scoring values 1.5→52.91 and 2.0→52.97 on seed 0; the λ threshold of 0.15 had zero effect, confirming most signal already comes from high-λ genes, so only α matters).

Round 1 showed the λ threshold at 0.15 is a no-op (scores identical to parent), meaning nearly all displacement mass comes from genes with λ≥0.15. The displacement cap at 90th percentile was catastrophically bad (removed too much signal). So the only remaining knob is α itself. α=1.5→52.91, α=2.0→52.97; the midpoint 1.75 should capture most of the cell_state gain of 2.0 with less covariation loss.

<<<<<<< SEARCH
    parser.add_argument("--alpha", type=float,
                        default=float(os.environ.get("VEC_ALPHA", "1.5")))
=======
    parser.add_argument("--alpha", type=float,
                        default=float(os.environ.get("VEC_ALPHA", "1.75")))
>>>>>>> REPLACE