Retire the Qwen AR pipeline

src/engine/metal/qwen_mtplx/policy.rs

@@ -53,7 +53,6 @@ pub(super) fn server_overrides(mtp: bool, get: impl Fn(&str) -> Option<String>)
         result.insert("MTPLX_SKIP_VERIFY_SNAPSHOT", "0");
     }
     for key in [
-        "MTPLX_AR_PIPELINE",
         "MTPLX_FAMILY_CAPTURE_COMMIT",
         "MTPLX_FUSED_HC_V3",
         "MTPLX_FUSED_GDN_INPROJ",
@@ -137,7 +136,6 @@ pub(super) fn turn_options(
             max_tokens: usize::MAX,
             seed: 0,
             stop_tokens: &[],
-            pipeline: flag("MTPLX_AR_PIPELINE"),
             sync_eval: !flag("MTPLX_ASYNC_AR")
                 || get("MTPLX_EVAL_AUDIT").is_some_and(|s| !s.is_empty()),
             return_hidden: false,
@@ -182,6 +180,42 @@ pub(super) fn turn_options(
     ))
 }
 
+#[test]
+fn retired_ar_pipeline_flag_cannot_change_server_or_turn_policy() {
+    let mut settings = crate::settings::effective_settings(
+        crate::model::ModelChoice::Qwen38FlashNext,
+        &crate::settings::GenerationPreferences::default(),
+        &crate::settings::RuntimePreferences::default(),
+        std::path::Path::new("/unused-policy-models"),
+    )
+    .unwrap()
+    .engine;
+    settings.speculative.glm_mtp = false;
+    for legacy in [None, Some("0"), Some("1")] {
+        let lookup = |key: &str| {
+            (key == "MTPLX_AR_PIPELINE")
+                .then_some(legacy)
+                .flatten()
+                .map(str::to_owned)
+        };
+        let overrides = server_overrides(false, lookup);
+        assert_eq!(overrides, server_overrides(false, |_| None));
+        let (options, _) = turn_options(&settings, |key| {
+            assert_ne!(key, "MTPLX_AR_PIPELINE", "retired option was read");
+            overrides
+                .get(key)
+                .map(|s| (*s).to_owned())
+                .or_else(|| lookup(key))
+        })
+        .unwrap();
+        let DecodeOptions::Ar(ar) = options.decode else {
+            panic!("AR request changed modes");
+        };
+        assert!(ar.sync_eval && ar.capture_final_state);
+        assert!(!ar.return_hidden);
+    }
+}
+
 #[test]
 #[ignore = "requires DS4_QWEN38_POLICY_REFERENCE from pinned MTPLX; CPU-only"]
 fn qwen_server_policy_matches_original_functions() {
@@ -198,6 +232,23 @@ fn qwen_server_policy_matches_original_functions() {
         let mtp = case["mtp"].as_bool().unwrap();
         let lookup = |key: &str| case["environment"][key].as_str().map(str::to_owned);
         let overrides = server_overrides(mtp, lookup);
+        if case["server_only"] == true {
+            // This scoped comparison owns AR retirement and the existing
+            // defaults. New upstream features are outside #104; retain the
+            // complete legacy policy comparison below for its own oracle.
+            assert!(
+                !case["overrides"]
+                    .as_object()
+                    .unwrap()
+                    .contains_key("MTPLX_AR_PIPELINE")
+            );
+            assert!(!overrides.contains_key("MTPLX_AR_PIPELINE"));
+            for (key, value) in &overrides {
+                assert_eq!(case["overrides"][key], *value, "default {key} case {count}");
+            }
+            count += 1;
+            continue;
+        }
         assert_eq!(
             serde_json::to_value(&overrides).unwrap(),
             case["overrides"],
@@ -253,7 +304,6 @@ fn qwen_server_policy_matches_original_functions() {
         match actual.decode {
             DecodeOptions::Ar(o) => {
                 assert!(!mtp);
-                assert_eq!(o.pipeline, expected["pipeline"]);
                 assert_eq!(o.sync_eval, expected["sync_eval"]);
             }
             DecodeOptions::Mtp(o) => {