## kernels.harness.run_retro()


Run every task through the staged escalation, in parallel.


Usage

``` python
kernels.harness.run_retro(
    tasks,
    *,
    stages=(SamplerSettings(),),
    gates=None,
    max_workers=None,
    executor="process",
    runner=run_task,
    progress=None
)
```


Stage 1 fits all tasks at `stages[0]`; each later stage re-fits only the tasks whose latest row fails `gates`. A task's final row is its LAST attempt (stamped `row["stage"]`), so escalation replaces, never duplicates. Rows come back in task order regardless of completion order.

`executor="serial"` runs everything inline in this process - for tests, debugging (real tracebacks), and platforms where spawning is unwanted. `runner` is injectable for the same reason. `progress` (if given) is called as `progress(row, stage, done_in_stage, total_in_stage)` from the parent as each fit lands.