chess-sim code: v2 corrections, precision metrics, jaw tolerance test

chess-sim/code/sim/push_dataset.py

@@ -63,7 +63,9 @@ NOTES = {
                             "policy hover (20%). Kept for reference; not used for training.",
     "baseline_recovery_v2": "Correction demonstrations, second attempt (`collect_recovery.py --style v2`): the "
                             "teacher lines up from where the arm is, rising only as far as needed, and goes "
-                            "straight down, with no pause. For fine-tuning the baseline together with baseline_v1.",
+                            "straight down, with no pause. Fine-tuning the baseline on these plus baseline_v1 "
+                            "lowered success (30% vs 51.7% on the same episodes: more pieces knocked over). "
+                            "Kept for reference; not used.",
 }