fix(miles): make shrink_engines post-sleep VRAM assert unconditional

[zhenyulincs]

Problem

RolloutManager.shrink_engines exposed an opt-in parameter for its Step-5 post-sleep VRAM leak check:

def shrink_engines(self, engine_indices, *, post_sleep_vram_threshold_gb: float | None = None):
    ...
    if post_sleep_vram_threshold_gb is not None:   # default None → skipped
        ... assert_post_sleep_vram_below_threshold ...

This made the handoff correctness check effectively dead: there is no static call site for shrink_engines in the repo (it is invoked dynamically by the external RLix coordinator via Ray RPC), and the parameter is keyword-only with default None, so no caller ever threaded a value into it — the Step-5 assert never executed.

Correction (vs. an earlier version of this description): the --miles-post-sleep-vram-threshold-gb arg does exist in arguments.py (default 1.0). An earlier draft claimed it "was never defined" — that was wrong (a grep with underscores missed the dashed argparse flag). The accurate statement is the narrower one above: the value was simply never passed into shrink_engines.

Why it should not be a parameter

A colocated GPU is time-shared — only one of train/infer holds VRAM at any moment. shrink_engines calls release_memory_occupation(tags=None) (full release of weights/KV/cuda_graph), so post-release residency must drop to near-zero or the next phase OOMs. That makes this assert a correctness floor for the GPU handoff, not a tunable knob. It should always run, with a fixed threshold.

Change

• Remove the post_sleep_vram_threshold_gb parameter from shrink_engines.
• Add module constant POST_SLEEP_VRAM_THRESHOLD_GB = 1.0 (GiB) with a comment explaining why it's a fixed floor.
• Run the Step-5 assert unconditionally (Anti-regression invariant refactor(miles): drop unused router engine metadata #8).
• Fix the stale sglang_engine.py docstring that referenced the (now-removed) arg.
• Remove the orphaned --miles-post-sleep-vram-threshold-gb argparse flag (arguments.py): nothing reads args.miles_post_sleep_vram_threshold_gb after the value is hardcoded, and its default of 1.0 already matched the constant.

Safety

• Removing the parameter breaks no caller: zero static call sites for shrink_engines in the repo, and the external coordinator could not have been passing it (the value was never wired in).
• Removing the arg breaks nothing: verified zero remaining references to miles_post_sleep_vram_threshold_gb (attribute) or --miles-post-sleep-vram-threshold-gb (flag) across the repo after the change.

Test plan

• ast.parse clean on all changed files
• No residual post_sleep_vram_threshold_gb parameter references
• No residual references to the removed arg (flag or attribute)
• RLix-mode shrink/expand smoke confirms the assert fires on a real release (and does not false-positive at ~0 GiB residency)