feat(context): improve budget validation and XML safety in ranking and Claude adapter

- Added stricter budget validation in ContextRanker with explicit error handling for invalid configurations.
- Introduced `_get_valid_token_count()` helper to validate and safeguard token counts.
- Enhanced XML escaping in Claude adapter to prevent injection risks from scores and unhandled content.

backend/app/services/context/adapters/claude.py

@@ -90,7 +90,9 @@ class ClaudeAdapter(ModelAdapter):
         elif context_type == ContextType.TOOL:
             return self._format_tool(contexts)
 
-        return "\n".join(c.content for c in contexts)
+        # Fallback for any unhandled context types - still escape content
+        # to prevent XML injection if new types are added without updating adapter
+        return "\n".join(self._escape_xml_content(c.content) for c in contexts)
 
     def _format_system(self, contexts: list[BaseContext]) -> str:
         """Format system contexts."""
@@ -119,7 +121,9 @@ class ClaudeAdapter(ModelAdapter):
             score = ctx.metadata.get("score", ctx.metadata.get("relevance_score", ""))
 
             if score:
-                parts.append(f'<document source="{source}" relevance="{score}">')
+                # Escape score to prevent XML injection via metadata
+                escaped_score = self._escape_xml(str(score))
+                parts.append(f'<document source="{source}" relevance="{escaped_score}">')
             else:
                 parts.append(f'<document source="{source}">')