Fix model metadata preservation with continuation_id (fixes #111) (#113)

* Fix model metadata preservation when using continuation_id

When continuing a conversation without specifying a model, the system now
correctly retrieves and uses the model from the previous assistant turn
instead of defaulting to DEFAULT_MODEL. This ensures model continuity
across conversation turns and fixes the metadata mismatch issue.

The fix:
- In reconstruct_thread_context(), check for previous assistant turns
- If no model is specified in the continuation request, use the model
  from the most recent assistant turn
- This preserves the model choice across conversation continuations

Added comprehensive tests to verify the fix handles:
- Single turn conversations
- Multiple turns with different models
- No previous assistant turns (falls back to DEFAULT_MODEL)
- Explicit model specification (overrides previous turn)
- Thread chain relationships

Fixes issue where continuation metadata would incorrectly report
'llama3.2' instead of the actual model used (e.g., 'deepseek-r1-8b')

* Update test to reference issue #111

* Refactor tests to call reconstruct_thread_context directly

Address Gemini Code Assist feedback by removing duplicated implementation
logic from tests. Tests now call the actual function with proper mocking
instead of reimplementing the model retrieval logic.

This improves maintainability and ensures tests validate actual behavior
rather than their own copy of the logic.

server.py

@@ -856,6 +856,16 @@ async def reconstruct_thread_context(arguments: dict[str, Any]) -> dict[str, Any
     # Create model context early to use for history building
     from utils.model_context import ModelContext
 
+    # Check if we should use the model from the previous conversation turn
+    model_from_args = arguments.get("model")
+    if not model_from_args and context.turns:
+        # Find the last assistant turn to get the model used
+        for turn in reversed(context.turns):
+            if turn.role == "assistant" and turn.model_name:
+                arguments["model"] = turn.model_name
+                logger.debug(f"[CONVERSATION_DEBUG] Using model from previous turn: {turn.model_name}")
+                break
+
     model_context = ModelContext.from_arguments(arguments)
 
     # Build conversation history with model-specific limits