# Tests run for this fix

Host: the atl build host, Python 3.14.4, pytest 9.0.2. These are development tests, not a
third-party audit.

| Command | Result |
|---|---|
| `python3 -m pytest tests/test_cain_mr_autoheal.py -q` | **12 passed** in 1.31 s |
| `python3 -m pytest tests/test_cain42_pbft_view_divergence.py tests/test_cain42_pbft_anti_entropy.py tests/test_cain42_pbft_evolution1.py -q` | **46 passed** in 31.99 s |
| `python3 -m pytest tests/test_cain42_pbft*.py tests/test_live_cluster_health.py -q` | **345 passed** in 294.18 s |

What the new tests cover:

- Every live cluster node has a deploy record that `pbft_alive()` can read; a cain-mr3 replica is
  recognised as alive when it answers PBFT and keeps up; a replica more than `LAG_OK` commits
  behind is still eligible for restart.
- A stale NEW_VIEW whose proposals are below `commit_index` causes a **step down** to the
  cluster's view instead of an endless re-broadcast; a NEW_VIEW without a certificate never
  triggers a step-down.

The four decision records at 06:47Z showed the cluster on `mr-2478e9c` with all four replicas in
lockstep; the first soak fault was handled with 5 expected check errors, 0 unexpected, and 0
divergences.
