fix(training): restore ground_tilt_penalty, add grounded_upright_fraction diagnostic

Round 4 changed three things at once and two of them cut upright pressure:
grounded_upright_reward went to 0 and ground_tilt_penalty was cut 2.5x,
while the new uprightness multiplier only pays below GROUND_HANDLING_HEIGHT
*and* while moving forward *and* facing the ball - a far narrower slice of
ticks than the penalty it was meant to replace. Net pressure fell and
upright_fraction fell with it (0.268 -> 0.239 -> 0.238, the lowest of any
round). Restore ground_tilt_penalty to 0.05 and change nothing else, so
this is a genuine single-variable test of multiplier plus full tilt
pressure.

The conjunctive mechanism itself held up: forward_motion_fraction reached
its best sustained value (0.242) without goal_rate sagging, ep_rew_mean
turned positive for the first time (+0.28), and eval win rate hit 49% with
no reward hacking.

Also adds grounded_upright_fraction: a diagnostic, deliberately ungated
metric measuring uprightness over real floor-contact ticks instead of
sub-3m ticks. upright_fraction has never exceeded 0.331 across four rounds
and ~560M steps without cheating, and its denominator is dominated by
ballistic transit (airborne_fraction ~0.45, mean_altitude ~4.4m) where
attitude is not meaningfully controllable - so it likely cannot measure
what the 0.45 floor was meant to capture. Re-baseline that floor from what
this reports rather than from another round of reshaping.

training/generation5.py

@@ -95,11 +95,33 @@ STANDING_ARGS = ["--ent-coef", "0.01", "--entropy-floor"]
 # forward-velocity-to-ball rises 0.06 -> 0.15 because multiplying by
 # uprightness cuts that term's expected per-tick value roughly 2-3x at
 # current behaviour; without the raise the approach incentive would quietly
-# weaken. ground-tilt-penalty drops 0.05 -> 0.02: it was the term most
-# likely to teach floor-avoidance, and with uprightness now positively paid
-# during play it only needs to remain a guardrail against travelling on the
-# roof rather than carry the whole posture signal. non-forward-penalty is
-# unchanged at 0.04 — it targets a specific behaviour and has not misfired.
+# weaken. non-forward-penalty is unchanged at 0.04 — it targets a specific
+# behaviour and has not misfired.
+#
+# Round 5 (2026-08-14): round 4 also dropped ground-tilt-penalty 0.05 ->
+# 0.02 on the theory that the multiplier could carry posture on its own.
+# That confounded the experiment — two of the three changes *reduced*
+# upright pressure at once (grounded_upright_reward to 0, tilt penalty cut
+# 2.5x) while the multiplier only pays below GROUND_HANDLING_HEIGHT *and*
+# while moving forward *and* facing the ball, i.e. a far narrower slice of
+# ticks than the penalty it replaced. Net pressure fell and so did
+# upright_fraction (0.268 -> 0.239 -> 0.238, the lowest of any round). The
+# conjunctive part worked though: forward_motion_fraction reached its best
+# sustained value (0.242) *without* goal_rate sagging, ep_rew_mean turned
+# positive for the first time (+0.28), and the eval win rate hit 49% with
+# no reward hacking. So round 5 restores ground-tilt-penalty to 0.05 and
+# changes nothing else — a genuine single-variable test of multiplier plus
+# full tilt pressure.
+#
+# Also added this round: grounded_upright_fraction, a *diagnostic, ungated*
+# telemetry signal measuring uprightness over real floor-contact ticks
+# rather than sub-3m ticks. upright_fraction has never exceeded 0.331
+# across four rounds and ~560M steps without the policy cheating, and its
+# denominator is dominated by ballistic transit (airborne_fraction ~0.45,
+# mean_altitude ~4.4m) where attitude is not meaningfully controllable —
+# so it likely cannot measure the behaviour the 0.45 floor was meant to
+# capture. Re-baseline that floor from what the new signal reports rather
+# than from another round of reshaping.
 HANDLING_REWARD_FLAGS = [
     "--velocity-to-ball-weight", "0.04",
     "--forward-velocity-to-ball-weight", "0.15",
@@ -109,7 +131,7 @@ HANDLING_REWARD_FLAGS = [
     "--goal-reward", "80",
     "--speed-reward-weight", "0.0",
     "--tilt-penalty", "0.0002",
-    "--ground-tilt-penalty", "0.02",
+    "--ground-tilt-penalty", "0.05",
     "--non-forward-penalty", "0.04",
     "--grounded-upright-reward", "0.0",
 ]