Retire the Qwen AR pipeline

tools/mtplx-policy-reference.py

@@ -6,6 +6,8 @@ import os
 
 parser = argparse.ArgumentParser(description=__doc__)
 parser.add_argument("model")
+parser.add_argument("--server-only", action="store_true",
+    help="Compare server defaults without the older boundary-cache option API")
 args = parser.parse_args()
 print("reference: importing policy functions without model loading", flush=True)
 from mtplx.server.openai import _server_runtime_env_overrides
@@ -47,6 +49,10 @@ for mtp, env in itertools.product((False, True), cases):
     overrides = _server_runtime_env_overrides(argparse.Namespace(
         model=args.model, generation_mode="mtp" if mtp else "ar"), None)
     os.environ.update(overrides)
+    if args.server_only:
+        print(json.dumps(dict(event="policy_reference", mtp=mtp, environment=env,
+            overrides=overrides, server_only=True)), flush=True)
+        continue
     # Native app selects bounded prefill; emulate that request-local policy.
     os.environ["MTPLX_SUSTAINED_PREFILL"] = "1"
     error = None
@@ -59,7 +65,7 @@ for mtp, env in itertools.product((False, True), cases):
                 lazy_kv=_lazy_snapshot_enabled(), store=g._store_on_prefill_env_enabled(),
                 store_min=g._store_on_prefill_min_suffix(), fused_max=g._small_suffix_fused_max(),
                 boundaries=g._gdn_boundary_capture_enabled(), boundary_max=g._gdn_boundary_max_count(),
-                boundary_tail=g._gdn_boundary_tail_interval(), pipeline=g._env_truthy("MTPLX_AR_PIPELINE"),
+                boundary_tail=g._gdn_boundary_tail_interval(),
                 sync_eval=not g._env_truthy("MTPLX_ASYNC_AR") or bool(os.environ.get("MTPLX_EVAL_AUDIT")),
                 capture=g._family_capture_commit_enabled() if mtp else None,
                 lazy_history=g._env_truthy("MTPLX_LAZY_MTP_HISTORY_APPEND"),