节点 n7

Looking at the results from rounds 0 and 1:
- Round 0: α=2.0 + displacement cap 1.5 → 52.924 (covariation 45.31, similar to baseline)
- Round 1: noise×2.0 + α=2.5 → 52.756 (covariation 43.07, worse)

The displacement cap in round 0 actually preserved covariation well (45.31 vs 45.94 baseline) but didn't improve cell_state enough. The issue is that α=2.0 with cap 1.5 gave cell_state 63.16 vs baseline 62.91 — barely any gain.

The baseline parent (α=1.5, no cap) scores 53.084. The key observation: α=1.5 is already near optimal. Both attempts to increase α or change shrinkage made things worse.

Let me try the opposite direction: **slightly reduce α to protect covariation more**, or try a different angle entirely.

Actually, looking more carefully at the data: the parent baseline with α=1.5 gives covariation 45.94. The original single-seed run showed 45.32. The 3-seed average is 45.94. 

A different idea: try α=1.25. This should sacrifice a little cell_state but gain back covariation. From the parent's grid: α=1.0 gave cell_state 59.25, covariation 46.69, total 52.14. α=1.5 gave 62.76, 45.32, total 52.91. Interpolating α=1.25: cell_state ~61, covariation ~46, total ~52.5. That's worse.

Another idea: keep α=1.5 but add a **small floor to the shrinkage** — i.e., ensure λ has a minimum value so that even weakly changing genes get a tiny push. No, that would hurt covariation.

Better idea: **reduce CAP_R from 3.0 to 2.0**. On X3, r=2.0 exactly, so this doesn't change X3 behavior at all. Not useful.

What about trying α=1.5 with a **weaker noise multiplier** (0.5×), meaning weaker shrinkage, more genes get displaced? This increases displacement for more genes. From the parent's data, more displacement helps cell_state but hurts covariation. With α=1.5 and weaker shrinkage, it's like effectively increasing α for marginal genes.

Actually, let me try: **noise multiplier = 0.5** (weaker shrinkage → more genes pass through) with **α = 1.5** (unchanged). The rationale: the EB shrinkage with the current noise level is quite aggressive (median λ≈0, only 8-24% of genes have λ>0.5). Weakening shrinkage lets more genes get displaced, which could improve de_recovery and cell_state. The risk is covariation, but since α stays at 1.5 (not increasing), the per-gene displacement magnitude stays moderate.

CHANGE: Weaken EB shrinkage by multiplying noise by 0.5 (more genes pass shrinkage threshold, more genes get displaced), keeping α=1.5; this should improve de_recovery and cell_state with moderate covariation cost since per-gene displacement magnitude is unchanged.

<<<<<<< SEARCH
            noise = 0.5 * (v1 + v2) * (1.0 / n1 + 1.0 / n2)
=======
            noise = 0.5 * (v1 + v2) * (1.0 / n1 + 1.0 / n2) * 0.5
>>>>>>> REPLACE

<<<<<<< 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", "1.5")))
    parser.add_argument("--noise-mult", type=float,
                        default=float(os.environ.get("VEC_NOISE_MULT", "0.5")))
>>>>>>> REPLACE

（只摘前 40 行）