fix(intern-decision-serve): register the checkpoint's inference module before executing it

Its dataclasses use postponed annotations and look their module up in sys.modules while the
class is built; importing by file path without registering it failed startup (closed).

services/intern-decision-serve/tests/test_engine.py

@@ -89,15 +89,17 @@ def reset_fake_torch():
 # A fake checkpoint: snapshots/<REVISION>/inference.py defining a DecisionEngine shaped like the real one.
 # ---------------------------------------------------------------------------------------------
 FAKE_INFERENCE = textwrap.dedent('''
+    from __future__ import annotations   # as in the real one: dataclasses then look the module up in sys.modules
     import json
+    from dataclasses import dataclass
     MODEL_NAME = "Intern-Decision-4B"
     EVENTS = []
     WARMUP_SHIFT = {"after_swap": 0.0}
     TOKEN_SKEW = {"n": 0}
 
+    @dataclass(frozen=True)          # the real inference.py defines one: needs sys.modules at exec time
     class Compiled:
-        def __init__(self, messages):
-            self.messages = messages
+        messages: list
 
     def validate_request(request):
         return {"state": request["state"], "questions": request["questions"]}