You are an agent, your current working directory is /app.

You can use the tools available to you to interact with the computer to assist the user in completing tasks.

# Repair incremental training after model persistence

The frozen scikit-learn source slice under `/app/vendor/scikit-learn` has a regression in a stateful CPU training workflow. The public harness trains an `MLPRegressor`, serializes and reloads it, then changes the target and performs repeated `partial_fit` calls with both Adam and momentum SGD.

Run:

```bash
python3 /app/run_regression.py
```

The current implementation advances optimizer bookkeeping but leaves the reloaded estimator's live weights and predictions unchanged. Diagnose the parameter ownership across the MLP fit loop and stochastic optimizers, then repair these existing production modules:

- `/app/vendor/scikit-learn/sklearn/neural_network/_multilayer_perceptron.py`
- `/app/vendor/scikit-learn/sklearn/neural_network/_stochastic_optimizers.py`

Acceptance requirements:

1. The optimizer update contract must apply gradients to the estimator's current coefficient and intercept arrays, including after serialization.
2. Existing Adam moments/step count and SGD momentum state must continue across reload and fine-tuning; do not recreate the optimizer on each incremental call.
3. Both public Adam and SGD cases must move materially closer to the changed target while preserving the fitted layer shapes and finite numeric state.
4. `python3 /app/run_regression.py` must complete and write `/app/results/finetune_report.json` with `output_schema_version` equal to `tbench.mlp_continuation.v1`.

Do not replace or edit `run_regression.py`, `runtime_loader.py`, `cases.json`, immutable vendored modules, or the license. Do not hard-code predictions, bypass serialization, change the workload, disable a solver, or install/download anything. The repair must be in both listed source modules and must remain compatible with both stochastic optimizers.