fix(training): cut grounded_upright_reward, restart stage-4 from Stage-3 foundation

grounded_upright_reward at 0.015 overshot: four force-retries pushed
upright_fraction from 0.265 to a plateauing 0.331, then the fifth jumped it
to 0.696 (55% over the 0.45 floor) while goal_rate collapsed 0.542->0.366
and forward_motion_fraction fell 0.244->0.184 (vertical_thrust_mean went
negative) - the policy learned to sit pinned upright and farm the bonus
instead of chasing the ball. It was sized "comparable to
time_penalty/ball_distance_penalty" but at 0.015/tick it was actually above
ball_distance_penalty's 0.01/tick worst case, so idling near the ball beat
playing. Cut to 0.004/tick (episode ceiling ~7.2, below
ball_distance_penalty's ~18 worst case). Delete the five blocked attempts
and reset generation5_state.json so the next run starts fresh from the
Stage-3 foundation rather than continuing from the farming checkpoint.

training/generation5.py

@@ -58,6 +58,23 @@ STANDING_ARGS = ["--ent-coef", "0.01", "--entropy-floor"]
 # reward is the positive counterpart: a bonus for genuine floor contact
 # (not just low altitude) while upright, so grounding well is the locally
 # profitable choice rather than merely the least-punished one.
+#
+# Round 2 (2026-08-11): grounded_upright_reward at 0.015 overshot. Four
+# force-retries pushed upright_fraction from 0.265 to a plateauing 0.331,
+# then the fifth jumped it to 0.696 (55% over the 0.45 floor) while
+# goal_rate collapsed 0.542->0.366 and forward_motion_fraction fell
+# 0.244->0.184 — the ship learned to sit pinned upright on the floor
+# (vertical_thrust_mean went negative) and farm the bonus instead of
+# playing. Root cause: 0.015/tick was actually *larger* than
+# ball_distance_penalty's worst case (0.01/tick), so idling near the ball
+# beat chasing it — not "comparable to time_penalty/ball_distance_penalty"
+# as originally sized. Cut to 0.004/tick (a full grounded episode now caps
+# at ~7.2, versus ball_distance_penalty's worst-case ~18 and a single goal's
+# 80) — enough to stop "avoid the floor" without being worth farming over
+# actually playing. Resets from the Stage-3 foundation again rather than
+# continuing from the farming checkpoint, same reasoning as the ground_tilt/
+# non_forward_penalty retune: don't resume a policy shaped by one reward
+# balance into a meaningfully different one.
 HANDLING_REWARD_FLAGS = [
     "--velocity-to-ball-weight", "0.04",
     "--forward-velocity-to-ball-weight", "0.06",
@@ -69,7 +86,7 @@ HANDLING_REWARD_FLAGS = [
     "--tilt-penalty", "0.0002",
     "--ground-tilt-penalty", "0.05",
     "--non-forward-penalty", "0.04",
-    "--grounded-upright-reward", "0.015",
+    "--grounded-upright-reward", "0.004",
 ]
 
 STAGES = [