fix(wan rollout)(5/n): drop the FlowGRPO negative-prompt fallback, use the engine default

[Rockdu]

Stack position: 5/n, on top of #170 (4/n).

What

• build_rollout_generate_payload no longer forces negative_prompt = " " whenever CFG is on,
  and build_rollout_sampling_params omits the key entirely when --diffusion-negative-prompt
  is unset. The engine then applies its own per-model default (rollout_api.py strips absent
  fields; for wan2.2 the default is the official 137-char Chinese negative prompt from
  wan/configs/shared_config.py).
• The SD3 recipes pass --diffusion-negative-prompt ' ' explicitly to keep their FlowGRPO
  alignment byte-for-byte.

Why

The " " fallback is an SD3/FlowGRPO convention that was silently imposed on every model family.
Model-specific defaults belong to the engine; miles should transmit what the user set and nothing
else. Measured on wan2.2 true inference (same prompt/seed/steps): the official negative prompt cuts
whole-video entropy ~20% and visibly cleans the first frame vs " ".

Parity-safe: the trainer recomputes from the engine-returned cond_kwargs, so both sides always see
the same negative embedding. Verified in production — model_output_mean_abs_diff = 0.0 sustained
after the switch (run wan22_16gpu_480p5f_ns16_n07_shift5).

Landing note

scripts/run_diffusion_grpo_cosmos3_pickscore_t2i_4gpu.py landed on main after this stack's base
and also relies on the fallback (CFG on, no explicit negative prompt). It needs the same one-line
--diffusion-negative-prompt ' ' when this stack merges; it does not exist on this branch.

Files

• miles/rollout/sglang_diffusion_rollout.py — drop the fallback; omit the key when unset
• scripts/run_diffusion_grpo_sd3_ocr_sglang.py, scripts/run_diffusion_grpo_pickscore_5gpu_flowgrpo_aligned.py
  — explicit ' ' to preserve behavior
• tests/fast/rollout/test_rollout_negative_prompt.py — unset ⇒ key absent; explicit ⇒ verbatim

Checklist

• Added/updated tests (test_rollout_negative_prompt.py, 2 passed, registered stage-a-cpu)
• isort/black --check clean against the repo config
• pre-commit run --all-files — not run locally (formatters verified individually)
• No new launch flags (--diffusion-negative-prompt already exists; its unset semantics change
  from " " to engine default — that is the fix)