fix: address PR review issues for auto-mode refactor

- agent-executor: move executeQuery into try block for proper heartbeat cleanup,
  re-parse tasks when edited plan is approved
- auto-loop-coordinator: handle feature execution failures with proper logging
  and failure tracking, support backward-compatible method signatures
- facade: delegate getActiveAutoLoopProjects/Worktrees to coordinator,
  always create own AutoLoopCoordinator (not shared), pass projectPath
  to approval methods and branchName to failure tracking
- global-service: document shared autoLoopCoordinator is for monitoring only
- execution-types: fix ExecuteFeatureFn type to match implementation
- feature-state-manager: use readJsonWithRecovery for loadFeature
- pipeline-orchestrator: add defensive null check and try/catch for
  merge response parsing
- plan-approval-service: use project-scoped keys to prevent cross-project
  collisions, maintain backward compatibility for featureId-only lookups

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

apps/server/src/services/feature-state-manager.ts

@@ -60,8 +60,12 @@ export class FeatureStateManager {
     const featurePath = path.join(featureDir, 'feature.json');
 
     try {
-      const data = (await secureFs.readFile(featurePath, 'utf-8')) as string;
-      return JSON.parse(data);
+      const result = await readJsonWithRecovery<Feature | null>(featurePath, null, {
+        maxBackups: DEFAULT_BACKUP_COUNT,
+        autoRestore: true,
+      });
+      logRecoveryWarning(result, `Feature ${featureId}`, logger);
+      return result.data;
     } catch {
       return null;
     }