fix: address review — remove redundant guard, add regression tests

- Drop the now-subsumed `normalized.startswith("qwen")` guard
  (token-prefix matching already covers the bare case)
- Add test_qwen_prefixed_alias_reasoning_detection covering:
  prefixed Qwen 3+ (al-/sn-), bare Qwen 3+, and Qwen 2.x exclusion

api/config.py

@@ -3316,11 +3316,12 @@ def _candidate_supports_reasoning(candidate: str) -> bool:
             if major >= 4 or (major == 3 and minor >= 7):
                 return True
         return False
-    if any(token.startswith("qwen") for token in tokens) or normalized.startswith("qwen"):
+    if any(token.startswith("qwen") for token in tokens):
         # Restrict to Qwen 3+ (exclude Qwen 2/2.5)
-        # Use token-prefix matching (position-independent) so prefixed ids like
+        # Token-prefix matching (position-independent) so prefixed ids like
         # "al-qwen3-8-max-preview" (tokens: ["al","qwen3","8",...]) still match,
-        # even though "qwen" is not a standalone token.
+        # even though "qwen" is not a standalone token. The bare
+        # normalized.startswith("qwen") case is subsumed by this check.
         match = re.search(r"qwen.*?(\d+)(?:\D+(\d+))?", normalized)
         if match:
             major = int(match.group(1))

tests/test_reasoning_effort_model_capabilities.py

@@ -514,3 +514,40 @@ def test_datestamped_claude3_not_reasoning_capable_heuristic():
             f"{model} must remain reasoning-capable"
         )
 
+
+def test_qwen_prefixed_alias_reasoning_detection():
+    """Prefixed Qwen IDs (e.g. New-API aliases) must still be detected.
+
+    Regression: "al-qwen3.8-max-preview" normalizes to tokens
+    ["al", "qwen3", "8", ...] where "qwen" is NOT a standalone token,
+    so the old exact-membership check silently failed.
+    """
+    # Prefixed Qwen 3+ → reasoning-capable
+    for model in (
+        "al-qwen3.8-max-preview",
+        "al-qwen3.7-max",
+        "al-qwen3.7-plus",
+        "al-qwen3.6-flash",
+        "sn-qwen3-235b-a22b",
+    ):
+        assert cfg._candidate_supports_reasoning(model) is True, (
+            f"{model} must be reasoning-capable (prefixed Qwen 3+)"
+        )
+    # Bare Qwen 3+ → still works
+    for model in (
+        "qwen3-235b-a22b",
+        "qwen3-32b",
+    ):
+        assert cfg._candidate_supports_reasoning(model) is True, (
+            f"{model} must be reasoning-capable (bare Qwen 3+)"
+        )
+    # Qwen 2.x → excluded regardless of prefix
+    for model in (
+        "al-qwen2.5-72b-instruct",
+        "qwen2.5-7b-instruct",
+        "qwen2-72b",
+    ):
+        assert cfg._candidate_supports_reasoning(model) is False, (
+            f"{model} must NOT be reasoning-capable (Qwen 2.x excluded)"
+        )
+