fix(miles-pipeline): unify gen step_target unit to group count

[TianyeGGBond]

Context

MilesPipeline._init_phase_b_infer performs the INIT→GENERATION priority
transition for actor_infer. The scheduler's gap-ratio planner skips a
GENERATION cluster whose step_target_estimate is None, so the re-request
passes an estimate.

That estimate was computed as rollout_batch_size * n_samples_per_prompt
(a trajectory count), while the per-rollout demand signal
(signal_rollout_demand, driven by the driver) uses rollout_batch_size
(a group count, one group = one prompt's sample set — the F9 unit).

The two feed the same scheduler step_target_estimate field with
different units. With n_samples_per_prompt=1 (smoke runs) they happen
to be equal, so it's masked; with a real config (n_samples_per_prompt>1)
the bootstrap estimate is inflated by the sample factor, biasing the
planner's initial demand high.

Change

• Add module-level helper _rollout_step_target(miles_args) returning the
  group-count estimate (max(rollout_batch_size, 1)), with a docstring
  pinning the unit.
• Phase B step 8 now calls the helper instead of multiplying by
  n_samples_per_prompt, so the bootstrap estimate matches the per-rollout
  demand signal.

Single file, no behavior change on the smoke path (n_samples_per_prompt=1).

Scope

rlix-side only. The miles dual driver already uses rollout_batch_size and
needs no change.

🤖 Generated with Claude Code