fix(agent): enhance SubTurn robustness and fix race conditions

Major improvements to SubTurn implementation:

**Fixes:**
- Channel close race condition (sync.Once)
- Semaphore blocking timeout (30s)
- Redundant context wrapping
- Memory accumulation (auto-truncate at 50 msgs)
- Channel draining on Finish()
- Missing depth limit logging
- Model validation

**Enhancements:**
- Comprehensive documentation (150+ lines)
- 11 new tests covering edge cases
- Improved error messages

All tests pass. Production-ready.

Related: #1316

pkg/agent/loop.go

@@ -300,10 +300,16 @@ func registerSharedTools(
 				spawnTool.SetAllowlistChecker(func(targetAgentID string) bool {
 					return registry.CanSpawnSubagent(currentAgentID, targetAgentID)
 				})
+
+				// Set SubTurnSpawner for direct sub-turn execution
+				spawner := NewSubTurnSpawner(al)
+				spawnTool.SetSpawner(spawner)
+
 				agent.Tools.Register(spawnTool)
-				
+
 				// Also register the synchronous subagent tool
 				subagentTool := tools.NewSubagentTool(subagentManager)
+				subagentTool.SetSpawner(spawner)
 				agent.Tools.Register(subagentTool)
 			} else {
 				logger.WarnCF("agent", "spawn tool requires subagent to be enabled", nil)
@@ -988,6 +994,7 @@ func (al *AgentLoop) runAgentLoop(
 			concurrencySem:       make(chan struct{}, 5), // maxConcurrentSubTurns
 		}
 		ctx = withTurnState(ctx, rootTS)
+		ctx = WithAgentLoop(ctx, al) // Inject AgentLoop for tool access
 		isRootTurn = true
 
 		// Register this root turn state so HardAbort can find it