AutoTPPR: minimise Test Top 20 DE MSE, do not maximise it

[GuoCheng24]

tasks/metric_config.json marks Gears / "Test Top 20 DE MSE" as true, which the loader reads as maximize:

"Gears": {
  "Test Top 20 DE MSE": true
}

The value comes from gears_model.test_metrics['mse_de'] (tasks/AutoTPPR/code/experiment.py:1409) — a mean squared error. The flag is read into MetricValue(..., maximize=maximize) at mcts_experiments_utils_claude.py:186-188 and then drives both node selection (improvement = new_metric - local_best_metric, :466) and reward (:516). So a worse test MSE is scored as an improvement and rewarded.

Worked through with the file's own values:

test MSE 0.20 -> 0.35 (worse)
  current config (maximize=true):  improvement = +0.15   scored as progress, rewarded
  with false:                      improvement = -0.15   scored as a regression

Two things make this look like a typo rather than intent:

1. Every sibling error metric in the same file uses false — IEEE39: {"val/PQ_Vm_rmse": false}, AutoMolecule3D: {"Forces MAE": false}, AutoScale: {"RMSE": false}.
2. The module's own fallback heuristic, one branch away at :193, infers the direction from the metric name and gets it right — default_maximize = not any(k in metric_lower for k in ["loss", "rmse", "error", "mse", "mae"]) returns False here. The explicit entry is strictly worse than having no entry at all.

One character.

Separately and not changed here: _get_metric_config at :38-46 has its return inside the for loop, so only the first metric of a multi-metric task config is ever reachable. Every task in the file currently has exactly one metric, so it is latent rather than active — mentioning it in case you want it fixed in the same pass.