节点 n74

I'll analyze this task carefully. This is a **draft** node for the `stochastic_bridge` direction (Schrödinger bridge / [SF]²M), but the current program is the `ot_moscot` seed. Let me think about what the task is actually asking.

Key observations:
1. The task says this is a **draft** node with direction `stochastic_bridge`
2. But the current program is `ot_moscot` (Waddington-OT coupling)
3. The instruction says "只改 `solution/run.py`" and the current program is the OT-based one
4. The direction library says for stochastic_bridge: "train a small flow with dual score heads in PCA (entropy-regularized pairing), express pairing uncertainty; extend the learned drift near endpoints in a controlled way"
5. The minimum implementation: "torchcfm `SchrodingerBridgeConditionalFlowMatcher`, fixed small σ, don't sweep many parameters; only use local drift at t=1 to push a small step, decode and add back residual"
6. The control comparison: "σ=0 (deterministic CFM) and 'only add same amount of noise, no drift'"

So the task is to **transform** this OT-based program into a stochastic bridge implementation. The current program uses moscot OT coupling; I need to replace that with a Schrödinger bridge / conditional flow matching approach using torchcfm.

Let me think about the design:

**Current program weaknesses** (from scores):
- cell_state: 50.07 (decent)
- covariation: 52.52 (decent)
- de_recovery: 51.38 (decent)
- direction: 49.14 (weakest - this is about whether cells move in the right direction)

The `direction` score being weakest suggests the displacement direction is not well-calibrated. The stochastic bridge approach with state-dependent drift could help here.

**Design for stochastic_bridge draft:**

The plan says:
- Train small flow with dual score heads in PCA (entropy-regularized pairing)
- Use torchcfm `SchrodingerBridgeConditionalFlowMatcher` with 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 "only add same amount of noise, no drift"

The key insight from the direction library: "bridge solves processes between known endpoints; bridge-outside continuation is an extra assumption, cannot directly extend training interval formula to t>1; diffusion amplifies erroneous states; gains only come from adding noise"

So the approach should be:
1. Use PCA embedding (same as current)
2. Train a Schrödinger bridge conditional flow matcher between prev and last stages in PCA space
3. At t=1, evaluate the learned velocity field to get a local drift for each cell
4. Push cells by a fraction of that drift (with controlled noise)
5. Decode back to gene space, add residual

（只摘前 40 行）