Add joint-sequence subtask training and collision-free FAST vocab mapping

- recipes/subtask_joint.yaml: paper-style single sequence (pi0.5 §IV-B) —
  the supervised subtask span gets text CE and conditions the FAST and
  flow losses in the same forward.
- joint_subtask_conditioning config flag rebuilds the same layout at
  inference: state on the task turn, generated subtask as a causal
  assistant turn (encode_prompt_with_targets + lang_causal_marks through
  sample_actions), in both the policy select_action path and the runtime
  adapter.
- fast_skip_tokens default 128 -> 1152 so FAST codes land below the <loc>
  range and never collide with VQA loc targets; _FAST_ACTION_VOCAB_SIZE
  tightened to the universal tokenizer's 1024 codes.
- Strip the trailing space from the 'Assistant:' generation prefill —
  SentencePiece folds the space into the first target token, so the
  space-suffixed prefill ended in a lone '▁' never seen in training.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/source/language_and_recipes.mdx

@@ -146,6 +146,12 @@ The renderer does not apply a tokenizer chat template. Policy processors decide
 Blend recipes select one weighted sub-recipe deterministically from the sample index.
 `recipes/subtask_mem.yaml` trains the compact core blend — high-level subtask prediction, low-level execution, and memory. `recipes/subtask_mem_vqa_speech.yaml` is the fuller variant that also adds VQA and spoken interjection responses.
 
+A message recipe with a supervised assistant turn on the `low_level` stream trains
+the π0.5 paper's joint sequence instead of a blend: the target span gets text CE
+while also conditioning the action losses in the same forward.
+`recipes/subtask_joint.yaml` is the provided example; pair it with
+`--policy.joint_subtask_conditioning=true` at inference.
+
 ## Graceful absence
 
 If both language columns are missing, `None`, or empty, `RenderMessagesStep` is a no-op.