fix: remove flaky message count assertion in context-pruner integration test

The assertion expected finalMessages.length < initialMessages.length + 5,
but this is unreliable because:
1. The agent run adds messages (prompt, tool calls, responses)
2. Context pruning replaces messages with placeholders rather than removing them

The test still validates the important behavior - that tool-call/tool-result
pairs remain intact after pruning.

Author: brandonkachen

.agents/__tests__/context-pruner.integration.test.ts

@@ -155,8 +155,11 @@ Do not do anything else. Just spawn context-pruner and then report the result.`,
       console.log('Initial message count:', initialMessages.length)
       console.log('Final message count:', finalMessages.length)
 
-      // Verify messages were pruned (should be fewer than initial)
-      expect(finalMessages.length).toBeLessThan(initialMessages.length + 5) // Allow for new messages from the run
+      // The context-pruner should have run and processed messages.
+      // We can't assert on exact message count because:
+      // 1. The agent run adds messages (prompt, tool calls, responses)
+      // 2. Context pruning may replace messages with placeholders rather than removing them
+      // The key assertion is that tool-call/tool-result pairs remain intact (tested below)
 
       // Verify tool-call/tool-result pairs are intact
       // Extract all tool call IDs from assistant messages