Fix audit hanging: add auto_apply param + allow_write for tool access

Root cause: claude agent without --dangerously-skip-permissions
hangs on tool permission prompts when stdin=DEVNULL.

Fixes:
- run_audit() now passes allow_write=True so agent can use
  Read/Bash tools without interactive permission prompts
- Added auto_apply param: False for API (result only),
  CLI confirms with user then applies manually
- API explicitly passes auto_apply=False
- Tests for auto_apply=True/False behavior

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

cli/main.py

@@ -613,6 +613,7 @@ def audit_backlog(ctx, project_id):
         return
 
     click.echo(f"Auditing {len(pending)} pending tasks for {project_id}...")
+    # First pass: get results only (no auto_apply yet)
     result = run_audit(conn, project_id)
 
     if not result["success"]:
@@ -643,9 +644,14 @@ def audit_backlog(ctx, project_id):
     if result.get("duration_seconds"):
         click.echo(f"Duration: {result['duration_seconds']}s")
 
+    # Apply: mark tasks as done after user confirmation
     if done and click.confirm(f"\nMark {len(done)} tasks as done?"):
         for item in done:
-            models.update_task(conn, item["id"], status="done")
+            tid = item.get("id")
+            if tid:
+                t = models.get_task(conn, tid)
+                if t and t["project_id"] == project_id and t["status"] == "pending":
+                    models.update_task(conn, tid, status="done")
         click.echo(f"Marked {len(done)} tasks as done.")