pi052: bump lm_head_lr_scale default to 5.0 (keep base LR at 2.5e-5)

The base optimizer LR (2.5e-5, cosine to 2.5e-6, 1k warmup, AdamW
(0.9, 0.95), wd 0.01, grad_clip 1.0) is the openpi/π0.5 setting used
for the RoboCasa leaderboard baselines and is well-validated for 3B-
class VLAs with a paligemma backbone. Leave it alone.

The one place pi052 needs to diverge from pi05 is the LM-head LR
multiplier:

  * pi05 has no text supervision -> head doesn't get gradients ->
    lm_head_lr_scale is moot, stays at 1.0.
  * pi052 always has text supervision via the recipe (subtask /
    memory / VQA). Under KI, the LM head only sees gradients on
    ~30-45% of the batch (the text-CE mask share). Under aggressive
    cosine decay the head drifts back toward PaliGemma's pretrained
    <loc> first-token bias, despite teacher-forced CE staying near 0.

5x is the documented fix (see PI05Config.lm_head_lr_scale docstring
and PI05Policy.get_optim_params, which is already wired to split the
LM head + tied embed_tokens into their own param group while sharing
the same cosine lambda). Flipping the default here lifts the fix from
opt-in to on-by-default for every pi052 run, with zero downside on
text-free recipes (head still gets no gradients to scale).

Other LR knobs reviewed and intentionally NOT changed:
  - optimizer_lr=2.5e-5: openpi-validated, matches leaderboard.
  - scheduler_warmup_steps=1000: standard for VLA finetuning.
  - scheduler_decay_steps=30000: auto-scales for short runs.
  - optimizer_betas=(0.9, 0.95): GPT/LLM convention, works for
    flow-matching + LM-CE.
  - optimizer_weight_decay=0.01, grad_clip=1.0: standard.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/lerobot/policies/pi052/configuration_pi052.py

@@ -163,6 +163,23 @@ class PI052Config(PI05Config):
     """If True, route every transformer layer through the KI
     attention path that blocks action→VLM gradient flow on K/V."""
 
+    # Learning-rate defaults --------------------------------------------
+    # pi052 inherits π0.5's openpi-validated optimizer config (peak LR
+    # 2.5e-5, cosine→2.5e-6, 1k warmup, AdamW (0.9, 0.95), wd=0.01,
+    # grad_clip=1.0). The only place pi052 needs to diverge from pi05
+    # is the LM-head LR multiplier: pi05 has no text supervision so the
+    # head doesn't get gradients; pi052 always has text supervision
+    # (subtask / memory / VQA) via the recipe, and under KI the LM head
+    # only sees gradients on ~30–45% of the batch (the text-CE mask
+    # share of the recipe). Under aggressive cosine decay this is too
+    # weak to keep the head pinned, so it drifts back toward PaliGemma's
+    # pretrained ``<loc>`` first-token bias. 5x is the documented fix
+    # (see ``PI05Config.lm_head_lr_scale`` docstring); the wiring is
+    # already in ``PI05Policy.get_optim_params`` — it splits the LM head
+    # + tied ``embed_tokens`` into their own param group while sharing
+    # the same cosine lambda, so the 5x ratio is preserved across decay.
+    lm_head_lr_scale: float = 5.0
+
     def __post_init__(self) -> None:
         super().__post_init__()
         # Backbone needs gradients flowing through the text head when