fix: handle missing </tool_call> closing tag in parse_tool_call

When using `</tool_call>` as the stop string, vLLM stops generation
BEFORE outputting the closing tag. This caused parse_tool_call() to
fail because the regex required both opening and closing tags.

Now handles both cases:
1. Full tag: <tool_call>...</tool_call>
2. Partial tag: <tool_call>... (when stop string is </tool_call>)

This was causing 0% reward in all Fleet task training runs because
valid tool calls were being rejected as "No tool call found".

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: 2 people

skyrl-train/integrations/fleet/env.py

@@ -78,11 +78,19 @@ def parse_tool_call(action: str) -> Optional[Dict[str, Any]]:
     - <tool_call>{"name": "...", "arguments": {...}}</tool_call>
     - <function_call>{"name": "...", "arguments": {...}}</function_call>
 
+    Also handles cases where the closing tag is missing (e.g., when </tool_call>
+    is used as the stop string and not included in the output).
+
     Returns dict with "name" and "arguments" keys, or None if not found.
     """
     # Try common tag formats
     for tag in ["tool_call", "function_call"]:
+        # First try with closing tag
         match = re.search(rf"<{tag}>(.*?)</{tag}>", action, re.DOTALL)
+        if not match:
+            # Try without closing tag (for when </tool_call> is the stop string)
+            # Match from opening tag to end of string or next special token
+            match = re.search(rf"<{tag}>(.*?)(?:<\||\Z)", action, re.DOTALL)
         if match:
             try:
                 parsed = json.loads(match.group(1).strip())