annotate: cap embedded-frame budget to fit VLM context (fix 32k overflow)

Switching the plan module to embedded frames (use_video_url=false)
exposed a context overflow: at frames_per_second=2.0 with the old
max_video_frames=128 default, a 480x640 episode embeds ~128 frames ≈
33-39k vision tokens, over the model's 32768 context — every plan call
died with 'Input length exceeds maximum context length' (HTTP 400),
crashing the whole annotation job.

The video_url path never hit this because the server downsampled; the
embedded path sends every sampled frame, so the frame count is a hard
token budget.

Fix:
  * config default max_video_frames 128 -> 32 (~8-10k vision tokens,
    comfortable headroom for the prompt + describe/verify passes).
    Frames are still sampled UNIFORMLY across the whole episode, so
    longer episodes are subsampled, not truncated — full temporal
    coverage preserved, just coarser density.
  * run_hf_job.py: frames_per_second 2.0 -> 1.0, explicit
    --plan.max_video_frames=32, with a comment explaining the token
    budget and the 'do not raise toward 128 with embedded frames' rule.

Only the plan module embeds the full episode; VQA (1 frame/tick) and
interjections (4-frame window) were never at risk.

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

examples/annotations/run_hf_job.py

@@ -58,10 +58,18 @@ CMD = (
     # handing the server a file:// clip. The embedded path is more
     # reliable: if clip extraction ever fails, the video_url path would
     # silently send NO video and the VLM would hallucinate subtasks from
-    # the task text alone. 2 fps gives dense visual grounding so the VLM
-    # labels what actually happens.
-    "--plan.frames_per_second=2.0 "
+    # the task text alone.
+    #
+    # CONTEXT BUDGET: with embedded frames, each frame is ~250-320 vision
+    # tokens. The model's context is 32768 (see --max-model-len). 32
+    # frames sampled uniformly across the episode (~8-10k tokens) fits
+    # comfortably alongside the prompt and the describe/verify passes.
+    # Do NOT raise max_video_frames toward 128 with embedded frames — that
+    # is ~33-39k tokens and overflows the context (BadRequestError 400,
+    # "Input length exceeds maximum context length").
     "--plan.use_video_url=false "
+    "--plan.frames_per_second=1.0 "
+    "--plan.max_video_frames=32 "
     # IMPORTANT for RoboCasa: the dataset's task string ("Navigate to the
     # stove", "Pick the mug...") is authoritative and is what eval uses.
     # ``derive_task_from_video=off`` keeps that canonical task driving

src/lerobot/annotations/steerable_pipeline/config.py

@@ -44,9 +44,19 @@ class PlanConfig:
     derive_task_from_video: str = "if_short"
     derive_task_min_words: int = 3
 
-    # Frame sampling for the subtask-decomposition prompt.
+    # Frame sampling for the subtask-decomposition prompt. Frames are
+    # sampled uniformly across the whole episode up to ``max_video_frames``
+    # (so longer episodes are subsampled, not truncated).
+    #
+    # ``max_video_frames`` is a HARD context-budget cap. With the embedded-
+    # frame path (use_video_url=false), every frame becomes ~250-320 vision
+    # tokens, so 128 frames ≈ 33-39k tokens — over a 32k-context VLM. 32
+    # frames (~8-10k tokens) leaves ample room for the prompt + the
+    # describe / verify passes. Raise only if your serving context is
+    # larger AND your episodes need finer temporal resolution; if you hit
+    # "Input length exceeds maximum context length", lower this.
     frames_per_second: float = 1.0
-    max_video_frames: int = 128
+    max_video_frames: int = 32
 
     min_subtask_seconds: float = 1.5
     plan_max_steps: int = 8