Tasks permanently stuck in upstream_failed after marking a failed task as success via API

### Apache Airflow version 3.1.8 ### If "Other Airflow 3 version" selected, which one? _No response_ ### What happened? When a task fails, the scheduler starts marking its downstream tasks as upstream_failed one layer at a time. If you mark the failed task as success via the API (PATCH) while the scheduler is still cascading, some downstream tasks get permanently stuck in upstream_failed and never run — even though the root cause (the failed task) is now successful. This happens because the scheduler and the API operate on separate database connections. The scheduler takes a snapshot of task states at the start of each loop and uses that snapshot for the entire loop. If the API changes task states after the snapshot was taken but before the scheduler finishes evaluating, the scheduler makes decisions based on stale/outdated data. It sees tasks as upstream_failed that have already been cleared, and marks the next layer of downstream tasks as upstream_failed too. The problem is permanent because upstream_failed is a terminal state — the scheduler never re-evaluates tasks in that state. **Example:** DAG: task_a >> task_b >> task_c 1. task_a fails 2. Scheduler marks task_b as upstream_failed 3. User marks task_a as success via the API — this also clears task_b back to a runnable state 4. But the scheduler was mid-loop with old data, still sees task_b as upstream_failed, and marks task_c as upstream_failed 5. task_c is now stuck forever — task_a is success, task_b runs fine, but task_c never runs ### What you think should happen instead? After marking a task as success via the API, all downstream tasks should eventually be re-evaluated correctly and run if their dependencies are met. A concurrent scheduler loop should not be able to permanently corrupt task states based on stale data. ### How to reproduce **DAG**: a linear chain like fail_task >> t0 >> t1 >> t2 with default trigger_rule=all_success. 1. Trigger the DAG — fail_task fails intentionally 2. The scheduler star…