节点 n72

Let me understand the task:

1. This is a draft node for `stochastic_bridge` (Schrödinger bridge / [SF]²M)
2. I need to write a new `solution/run.py` that implements a Schrödinger bridge approach
3. The current program is `ot_moscot` with score 48.8657
4. I need to produce a SEARCH/REPLACE block for run.py + one METHOD block

Key constraints:
- View-agnostic: no hardcoded stage names, file names, or cell counts
- Must handle both 1-input (proxy) and 2-input (final) cases
- Deterministic with --seed
- No held-out stage data
- torchcfm is available (SchrodingerBridgeConditionalFlowMatcher)
- EXECUTION.json says gpu: false, so CPU only
- Must keep `solution/g37_common.py` as read-only import
- Output genes follow genes.txt panel
- Cells in [min_cells, max_cells]

The method card says:
- Train small flow + score dual-head in PCA (entropic-regularized pairing)
- Use drift learned near t=1 for a limited extrapolation step
- torchcfm SchrodingerBridgeConditionalFlowMatcher, fixed small σ
- Only use local drift at t=1 to push a small step, decode and add back residual
- Control: σ=0 (deterministic CFM) and "noise only, no drift"

Key insight from the experiment table:
- Best scores ~50-53, with cell_state being the weakest dimension for top performers
- The current ot_moscot gets 50.65 on X3 (cell_state 50.07, covariation 52.52, de_recovery 51.38, direction 49.14)
- direction is weakest here

Strategy for stochastic_bridge:
1. Use two input stages (prev, last) to train a Schrödinger bridge in PCA space
2. At inference, use the learned drift at t=1 to extrapolate beyond the last stage
3. Add controlled noise (σ > 0) to maintain spread
4. Fall back to growth-weighted copy for single-input views

Let me design the implementation:

1. Embed both stages in PCA (reuse g37_common.embed)
2. Train SchrodingerBridgeConditionalFlowMatcher on (Z_prev, Z_last) pairs

（只摘前 40 行）