fix(cache): drop ttl=1h on Portal Qwen — Alibaba upstream is 5m-only (#24702)

PR #24151 routed Portal Qwen (qwen3.6-plus) through the prefix_and_2
long-lived cache layout, attaching {"type":"ephemeral","ttl":"1h"}
markers to the tools[-1] entry and the stable system-prefix block.
That layout works for Portal Claude because Anthropic / OpenRouter on
Anthropic routes honour 1h TTL — but Portal Qwen ultimately proxies to
Alibaba DashScope, which documents a single "ephemeral" TTL of 5
minutes on its Context Cache. The ttl="1h" qualifier is silently
dropped upstream, so the two highest-value breakpoints (tools array +
system prefix) never land. Only the rolling-window 5m markers on the
last 2 messages cache, which matches the observed ~25% read rate.

Fix: keep Portal Qwen on cache_control via _anthropic_prompt_cache_policy
returning (True, False), but drop it from _supports_long_lived_anthropic_cache
so it rides the standard system_and_3 5m layout (system + last 3 messages,
all at 5m). Same 4 breakpoints, all in a TTL the upstream actually honours.

Refs: https://www.alibabacloud.com/help/en/model-studio/context-cache
      https://openrouter.ai/docs/features/prompt-caching (Alibaba Qwen
      section: "TTL: 5 minutes")

- _supports_long_lived_anthropic_cache: Portal scope narrowed back to Claude
- tests: flip the two qwen long-lived expectations to False, retitle
  non_claude_non_qwen_rejected -> non_claude_rejected

run_agent.py

@@ -3619,12 +3619,19 @@ class AIAgent:
         is_claude = "claude" in model_lower
         is_nous_portal = "nousresearch" in eff_base_url.lower()
 
-        # Nous Portal: Claude AND Qwen both get long-lived caching.
-        # Portal proxies to OpenRouter with identical cache_control
-        # semantics; any model on Portal that accepts envelope-layout
-        # markers via _anthropic_prompt_cache_policy also benefits from
-        # the documented 1h cross-session TTL.
-        if is_nous_portal and (is_claude or "qwen" in model_lower):
+        # Nous Portal Claude rides the 1h prefix_and_2 layout (Portal
+        # proxies to OpenRouter, which honours ttl=1h on Anthropic
+        # routes).  Qwen does NOT — Alibaba DashScope (the upstream for
+        # all Qwen routes, including Portal -> OpenRouter -> Alibaba)
+        # documents a single ``ephemeral`` TTL of 5 minutes; ttl="1h"
+        # on Qwen markers is silently ignored upstream, so the
+        # high-value tools[-1] + system-prefix breakpoints never land
+        # and only the 5m rolling-window markers on the last 2 messages
+        # get cached.  Portal Qwen still gets cache_control via
+        # _anthropic_prompt_cache_policy returning (True, False) — it
+        # just rides the standard system_and_3 5m layout instead of the
+        # mismatched prefix_and_2 1h layout.
+        if is_nous_portal and is_claude:
             return True
 
         if not is_claude:

tests/run_agent/test_anthropic_prompt_cache_policy.py

@@ -372,29 +372,36 @@ class TestSupportsLongLivedAnthropicCache:
         )
         assert agent._supports_long_lived_anthropic_cache() is True
 
-    def test_nous_portal_qwen_supported(self):
-        # Portal Qwen rides the same OpenRouter-equivalent transport as
-        # Portal Claude; long-lived (1h cross-session) cache_control
-        # markers apply identically.
+    def test_nous_portal_qwen_NOT_long_lived(self):
+        # Portal Qwen still gets cache_control markers via the standard
+        # system_and_3 5m layout (see _anthropic_prompt_cache_policy
+        # tests above), but it must NOT ride the prefix_and_2 1h layout.
+        # Alibaba DashScope (the upstream for every Qwen route, incl.
+        # Portal -> OpenRouter -> Alibaba) only supports a single
+        # ``ephemeral`` TTL of 5 minutes; ttl="1h" markers are silently
+        # ignored, so the high-value tools[-1] + system-prefix
+        # breakpoints don't land. Stay on system_and_3 instead.
         agent = _make_agent(
             provider="nous",
             base_url="https://inference-api.nousresearch.com/v1",
             api_mode="chat_completions",
             model="qwen3.6-plus",
         )
-        assert agent._supports_long_lived_anthropic_cache() is True
+        assert agent._supports_long_lived_anthropic_cache() is False
 
-    def test_nous_portal_qwen_vendored_slug_supported(self):
+    def test_nous_portal_qwen_vendored_slug_NOT_long_lived(self):
         agent = _make_agent(
             provider="nous",
             base_url="https://inference-api.nousresearch.com/v1",
             api_mode="chat_completions",
             model="qwen/qwen3.6-plus",
         )
-        assert agent._supports_long_lived_anthropic_cache() is True
+        assert agent._supports_long_lived_anthropic_cache() is False
 
-    def test_nous_portal_non_claude_non_qwen_rejected(self):
-        # Portal long-lived cache scope mirrors policy: Claude or Qwen only.
+    def test_nous_portal_non_claude_rejected(self):
+        # Portal long-lived cache scope is now Claude-only. Qwen
+        # rejection is covered by the dedicated tests above; this
+        # covers everything else (gpt, etc.).
         agent = _make_agent(
             provider="nous",
             base_url="https://inference-api.nousresearch.com/v1",