Replace vision hotswap with native multimodal profiles

dev/fake-profile-dir/override.conf

@@ -1 +1 @@
-ExecStart=/usr/bin/mock-llama --ctx-size 73728
+ExecStart=/usr/bin/mock-llama --ctx-size 76800

dev/fake-profile-dir/profile-fast.conf.disabled

@@ -1 +1 @@
-ExecStart=/usr/bin/mock-llama --ctx-size 73728
+ExecStart=/usr/bin/mock-llama --ctx-size 76800

dev/mock_upstream.py

@@ -25,7 +25,7 @@ def current_ctx() -> int:
                     return int(line.split("--ctx-size")[1].split()[0])
     except (OSError, ValueError, IndexError):
         pass
-    return 73728
+    return 76800
 
 
 class Handler(BaseHTTPRequestHandler):

dev/test_local.sh

@@ -124,7 +124,7 @@ import json,sys
 d=json.load(sys.stdin)
 ids={m["id"]:m for m in d["data"]}
 assert set(ids)=={"qwen-fast","qwen-medium","qwen-long"}, ids
-assert ids["qwen-fast"]["context_length"]==73728
+assert ids["qwen-fast"]["context_length"]==76800
 assert ids["qwen-medium"]["context_length"]==94208
 assert ids["qwen-long"]["context_length"]==131072
 ' && ok "drei virtuelle Modelle mit korrekten Context Windows" || bad "/v1/models"
@@ -138,7 +138,7 @@ import json,sys
 d=json.load(sys.stdin)
 assert d["current_profile"]=="fast", d
 assert d["upstream"]["reachable"] is True, d
-assert d["upstream"]["ctx"]==73728, d
+assert d["upstream"]["ctx"]==76800, d
 ' && ok "Status zeigt Profil fast + erreichbares Upstream" || bad "/status"
 
 # --- 3. Normales Forwarding (non-streaming) -------------------------------------
@@ -199,7 +199,7 @@ echo "$RESP" | python3 -m json.tool
 echo "$RESP" | python3 -c '
 import json,sys
 d=json.load(sys.stdin)
-assert d["profile"]=="fast" and d["context_length"]==73728, d
+assert d["profile"]=="fast" and d["context_length"]==76800, d
 ' && ok "Profil fast wieder aktiv" || bad "Profilwechsel fast"
 
 # --- 8. Virtuelles Modell triggert Profilwechsel ----------------------------------------
@@ -270,7 +270,7 @@ echo "$RESP" | python3 -m json.tool
 echo "$RESP" | python3 -c '
 import json,sys
 d=json.load(sys.stdin)
-assert d["profile"]=="fast" and d["model"]=="mock-model-73728", d
+assert d["profile"]=="fast" and d["model"]=="mock-model-76800", d
 ' && ok "Recovery: /fast wartet auf Upstream, dann Erfolg" || bad "Recovery"
 
 # --- 11. Bildgenerierung (Mock-Worker) -------------------------------------------------

dev/test_router_support.py

@@ -23,7 +23,11 @@ from router_support import (  # noqa: E402
     enforce_artifact_retention,
     load_profile_registry,
 )
-from ai_profile_router import _normalize_vision_image  # noqa: E402
+from ai_profile_router import (  # noqa: E402
+    _normalize_chat_image,
+    _normalize_chat_images,
+    _request_has_image,
+)
 
 
 class AuthPolicyTests(unittest.TestCase):
@@ -47,9 +51,9 @@ class RuntimeStoreTests(unittest.TestCase):
     def test_atomic_roundtrip_and_delete(self) -> None:
         with tempfile.TemporaryDirectory() as temp:
             store = RuntimeStore(str(Path(temp) / "state.json"))
-            store.save(worker="vision", worker_pid=123, last_profile="fast")
+            store.save(worker="image", worker_pid=123, last_profile="fast")
             self.assertEqual(store.load()["worker_pid"], 123)
-            store.clear_worker("vision")
+            store.clear_worker("image")
             state = store.load()
             self.assertNotIn("worker", state)
             self.assertEqual(state["last_profile"], "fast")
@@ -75,18 +79,38 @@ class ProfileRegistryTests(unittest.TestCase):
             load_profile_registry("/definitely/missing/profiles.json")
 
 
-class VisionInputTests(unittest.TestCase):
+class ChatImageInputTests(unittest.TestCase):
     def test_small_png_data_url_is_accepted(self) -> None:
         value = "data:image/png;base64,iVBORw0KGgo="
-        self.assertEqual(_normalize_vision_image(value), value)
+        self.assertEqual(_normalize_chat_image(value), value)
 
     def test_remote_url_is_denied_by_default(self) -> None:
         with self.assertRaisesRegex(ValueError, "deaktiviert"):
-            _normalize_vision_image("https://example.com/private.png")
+            _normalize_chat_image("https://example.com/private.png")
 
     def test_invalid_base64_is_rejected(self) -> None:
         with self.assertRaisesRegex(ValueError, "Base64"):
-            _normalize_vision_image("data:image/png;base64,not!base64")
+            _normalize_chat_image("data:image/png;base64,not!base64")
+
+    def test_multimodal_part_is_retained_for_direct_forwarding(self) -> None:
+        value = "data:image/png;base64,iVBORw0KGgo="
+        request = {
+            "model": "qwen-fast",
+            "messages": [{
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "Was ist zu sehen?"},
+                    {"type": "image_url", "image_url": {"url": value}},
+                ],
+            }],
+        }
+        self.assertTrue(_request_has_image(request))
+        normalized = _normalize_chat_images(request)
+        self.assertEqual(
+            normalized["messages"][0]["content"][1]["image_url"]["url"],
+            value,
+        )
+        self.assertEqual(request, normalized)
 
 
 class RetentionTests(unittest.TestCase):