test(llm): cover llama.cpp backend integration paths

src/gaia/governance/checkpoint_bridge.py

@@ -11,10 +11,13 @@
 
 from threading import Lock
 
+from typing import cast
+
 from .exceptions import CheckpointNotFoundError, InvalidResolutionError
 from .schemas import (
     CheckpointRecord,
     CheckpointResolution,
+    CheckpointStatus,
     GovernanceDecision,
     TransitionOutcome,
     WorkflowTransition,
@@ -90,7 +93,7 @@ def resolve_checkpoint(
                 checkpoint_id=current.checkpoint_id,
                 workflow_id=current.workflow_id,
                 transition_id=current.transition_id,
-                status=status,
+                status=cast(CheckpointStatus, status),
                 created_at=current.created_at,
                 decision_context={
                     **current.decision_context,

src/gaia/llm/factory.py

@@ -67,4 +67,4 @@ def create_client(
     module = importlib.import_module(module_path)
     provider_class = getattr(module, class_name)
 
-    return provider_class(**kwargs)
+    return provider_class(**kwargs)  # type: ignore[no-any-return]

src/gaia/llm/providers/claude.py

@@ -74,7 +74,7 @@ def chat(
         response = self._client.messages.create(**params)
         if stream:
             return self._handle_stream(response)
-        return response.content[0].text
+        return response.content[0].text  # type: ignore[union-attr]
 
     # embed() inherited from ABC - raises NotSupportedError
 
@@ -99,7 +99,8 @@ def vision(self, images: list[bytes], prompt: str, **kwargs) -> str:
                 ],
             }
         ]
-        return self.chat(messages, **kwargs)
+        result = self.chat(messages, **kwargs)
+        return result if isinstance(result, str) else "".join(result)
 
     # get_performance_stats() inherited from ABC - raises NotSupportedError
     # load_model() inherited from ABC - raises NotSupportedError

src/gaia/llm/providers/lemonade.py

@@ -57,13 +57,13 @@ class LemonadeContextOverflowError(LemonadeError):
 
     ``retryable`` is dynamic — set in ``_classify_lemonade_response`` based
     on the reported ``n_ctx``. When n_ctx is smaller than GAIA's expected
-    32K, the model was loaded with the wrong ctx_size; reloading via the
-    pre-flight helper will fix it, so we mark retryable so the chat layer
-    auto-recovers. When n_ctx is already at full size, this is a genuine
-    "conversation too big" situation and retry won't help.
+    65536 (64K), the model was loaded with the wrong ctx_size; reloading
+    via the pre-flight helper will fix it, so we mark retryable so the
+    chat layer auto-recovers. When n_ctx is already at full size, this is
+    a genuine "conversation too big" situation and retry won't help.
     """
 
-    retryable = False  # default; set True dynamically when n_ctx < 32K
+    retryable = False  # default; set True dynamically when n_ctx < 65536
     user_message = (
         "This conversation got too long for the model's context window. "
         "Start a fresh task to keep going."

src/gaia/llm/providers/openai_provider.py

@@ -60,7 +60,7 @@ def chat(
         )
         if stream:
             return self._handle_stream(response)
-        return response.choices[0].message.content
+        return response.choices[0].message.content  # type: ignore[union-attr]
 
     def embed(
         self, texts: list[str], model: str = "text-embedding-3-small", **kwargs