chore(typing): defer annotations in eager modules

gmlx/__main__.py

@@ -2,6 +2,8 @@
 script). A lazy import keeps ``-m gmlx <verb>`` light; this is also the form the
 background / launchd re-exec uses, so it must route to the umbrella (note that
 ``-m gmlx.commands.cli`` runs ``run``'s ``main`` instead)."""
+from __future__ import annotations
+
 import sys
 
 

gmlx/cache/kv_policy.py

@@ -6,6 +6,8 @@
 logging, and memory pricing all read the same object.
 """
 
+from __future__ import annotations
+
 from dataclasses import dataclass
 
 VALID_BITS = (2, 3, 4, 6, 8)

gmlx/serve/batch_rows.py

@@ -16,6 +16,8 @@
 of the batch as it currently is.
 """
 
+from __future__ import annotations
+
 
 def batch_rows(gen) -> int:
     """Active rows in ``gen._generation_batch``; 0 when there is none.

gmlx/serve/kv_policy.py

@@ -6,6 +6,8 @@
 object. An error verdict fails residency with the reason.
 """
 
+from __future__ import annotations
+
 import logging
 import math
 import os

gmlx/serve/step_timing.py

@@ -18,6 +18,8 @@
 (above the batch generator). No-op unless the env var is set.
 """
 
+from __future__ import annotations
+
 import atexit
 import os
 import sys

gmlx/spec/helpers.py

@@ -9,6 +9,8 @@
 Logic is a faithful copy (acceptance must stay token-identical to the validated
 path); keep it in sync when mlx-vlm's round changes in a way we want to track.
 """
+from __future__ import annotations
+
 import logging
 import os
 from dataclasses import dataclass