fix(cli): guard miner score _drain_logs against Warning-state crash

[statxc]

Summary

gitt miner score --log-level warning crashed at shutdown with
statemachine.exceptions.TransitionNotAllowed: Can't disable_logging when in Warning.
This guards the _drain_logs cleanup path so every value Click already accepts in
--log-level completes cleanly.

Closes #1373

Root cause

bittensor's LoggingMachine defines disable_logging transitions from
Trace/Debug/Default/Disabled/Info but not from Warning — the only
shutdown transition out of Warning is disable_warning = Warning.to(Default)
(bittensor/utils/btlogging/loggingmachine.py:127-137). Any caller that lands in
Warning and then invokes bt.logging.off() raises TransitionNotAllowed.

gitt miner score exposes warning via click.Choice([...]), calls
_apply_log_level('warning') → bt.logging.set_warning(), then unconditionally
calls bt.logging.off() in _drain_logs(). Deterministic crash, not flaky.

Fix

Single guarded transition in _drain_logs(): if the LoggingMachine is in
Warning, step down to Default via set_warning(on=False) first, then call
off() (which is defined from Default). This preserves the helper's existing
contract — before_disable_logging sets all stdlib loggers to CRITICAL — that
a naked skip-when-Warning would silently break.

State check follows bittensor's own internal idiom
(set_debug/set_trace/set_info/set_warning all compare
current_state_value before transitioning) and avoids using exceptions for
control flow.

A/B reproduction

Before:
info exit=0
debug exit=0
trace exit=0
warning exit=1 ← TransitionNotAllowed in _drain_logs

After:
info exit=0
debug exit=0
trace exit=0
warning exit=0

Tests

New TestDrainLogsAcrossAdvertisedLevels in tests/cli/test_miner_score.py:

• Parametrized over all four advertised levels (warning/info/debug/
  trace) — pins the regression and prevents drift if Click's --log-level
  choices ever expand.
• Disabled.to(Disabled) idempotency — a second drain after a successful
  drain must not raise.
• Warning → Disabled contract — proves the fix doesn't just sidestep the
  crash; the machine actually ends in Disabled so logging is genuinely off.

A restore_bt_logging_state fixture snapshots current_state_value before each
test and restores it on teardown — the LoggingMachine is a process-wide
singleton and would otherwise leak into the 867 other tests.

Test plan

• uv run --extra dev pytest tests/cli/test_miner_score.py -v → 21 passed
  (6 new + 15 existing)
• uv run --extra dev pytest tests/ -x -q → 873 passed
• uv run ruff check . / uv run ruff format --check . → clean
• uv run --extra dev pyright gittensor tests → 0 errors
• uv run --extra dev vulture → no findings
• Manual repro: A/B across all four --log-level values, all exit 0

[anderdc]

The guard in gittensor/cli/miner_commands/score.py is the fix and it's correct. The 50+ lines added to tests/cli/test_miner_score.py — the parametrized test class, the idempotency and lands-in-Disabled cases, and the process-wide restore_bt_logging_state singleton fixture — are disproportionate for a two-line guard. Keep: the score.py guard. Drop: the additions to tests/cli/test_miner_score.py.

[statxc]

Hi, @anderdc, I've removed. Thanks for the review.

[statxc]

Thanks @anderdc