[Bug] `gitt miner score --json --log-level warning` crashes in `_drain_logs()` — `bt.logging.off()` raises `TransitionNotAllowed` from `Warning` state

[taillred]

Scope

One-line guard in _drain_logs() so the cleanup path is tolerant of every --log-level value Click already accepts. Matches the reachable-crash acceptance shape of PR #624.

Description

gitt miner score advertises warning in click.Choice(['warning', 'info', 'debug', 'trace']) (score.py:255), but using that value crashes at shutdown. The command calls _drain_logs() (score.py:325), which unconditionally calls bt.logging.off() (score.py:235). When bittensor's LoggingMachine is in the Warning state, disable_logging is not a defined transition and the command exits with a traceback.

This is the only advertised --log-level value that crashes. info, debug, and trace all complete cleanly.

Bittensor library evidence (deterministic, not flaky)

From bittensor/utils/btlogging/loggingmachine.py:131-137:

disable_logging = (
    Trace.to(Disabled)
    | Debug.to(Disabled)
    | Default.to(Disabled)
    | Disabled.to(Disabled)
    | Info.to(Disabled)
)

Warning is intentionally omitted — bittensor's only Warning shutdown transition is disable_warning = Warning.to(Default) (line 127). Any caller that lands in Warning and then invokes bt.logging.off() raises TransitionNotAllowed. This is a stable bittensor API behavior, not a transient race.

Steps to Reproduce

From the repository root:

uv run --extra dev gitt miner score \
  --pat ghp_invalid_token_for_repro \
  --json \
  --log-level warning

Exit code 1, traceback ends in:

File ".../gittensor/cli/miner_commands/score.py", line 325, in score_command
    _drain_logs()
File ".../gittensor/cli/miner_commands/score.py", line 235, in _drain_logs
    bt.logging.off()
  ...
statemachine.exceptions.TransitionNotAllowed: Can't disable_logging when in Warning.

A/B across all four advertised --log-level values:

for lvl in info debug trace warning; do
  uv run --extra dev gitt miner score --pat ghp_invalid_token_for_repro \
    --json --log-level "$lvl" >/tmp/out 2>/tmp/err
  echo "$lvl exit=$?"
done

Output:

info exit=0
debug exit=0
trace exit=0
warning exit=1

Expected Behavior

Every value Click already accepts in --log-level should complete without a state-machine crash. The command should emit its JSON result or a normal CLI error envelope; it should not raise out of the logging-shutdown path.

Actual Behavior

--log-level warning exits 1 with statemachine.exceptions.TransitionNotAllowed: Can't disable_logging when in Warning.. All other advertised levels complete cleanly. The crash is independent of the scoring pipeline — _apply_log_level('warning') puts bittensor's LoggingMachine into the Warning state, and _drain_logs()'s unconditional bt.logging.off() then fires a transition bittensor does not define.

Proposed Fix

One-line state check in _drain_logs(). Matches bittensor's own internal pattern of comparing current_state_value before transitioning (see set_debug / set_trace / set_info in loggingmachine.py):

 def _drain_logs() -> None:
     ...
     import bittensor as bt

-    bt.logging.off()
+    # bittensor's LoggingMachine has no `disable_logging` transition from the
+    # Warning state (only `disable_warning = Warning.to(Default)` exists).
+    # info/debug/trace all permit the transition, so guard only Warning.
+    if bt.logging.current_state_value != 'Warning':
+        bt.logging.off()
     queue = bt.logging.get_queue()
     ...

State check is preferred over try/except TransitionNotAllowed so we don't use exceptions for normal control flow.

Regression test

Hermetic — doesn't need the full scoring pipeline. Assert that draining after each advertised log level is a no-op exception-wise and still reaches the flush path:

import pytest
import bittensor as bt
from gittensor.cli.miner_commands.score import _apply_log_level, _drain_logs

@pytest.mark.parametrize('level', ['warning', 'info', 'debug', 'trace'])
def test_drain_logs_tolerates_every_advertised_level(level):
    _apply_log_level(level)
    _drain_logs()  # must not raise
    # Idempotent: a second drain in the now-Disabled state must also not raise.
    _drain_logs()

The first call covers the Warning → Disabled crash this issue reports. The second covers Disabled → Disabled, which bittensor does allow but is worth pinning.

Affected Files

• gittensor/cli/miner_commands/score.py — _drain_logs (lines 223–243), _apply_log_level (lines 208–220).
• New test file or addition to tests/cli/test_miner_score.py.

Prior art / non-duplication

• PR fix: guard decode_packed_contract_storage against short buffers #624 (merged 2026-04-20): "fix: guard decode_packed_contract_storage against short buffers." This issue matches that acceptance shape — a narrow defensive guard for a reachable, demonstrated crash. Different module, same gate-only-on-real-crash pattern.
• PR feat: Be able to run scoring pipeline locally #947 (merged 2026-05-09): "feat: Be able to run scoring pipeline locally." Introduced the gitt miner score command and _drain_logs helper. This issue is a follow-up gap in that command.
• PR Standardize --json behavior #1152 (merged 2026-05-12): "Standardize --json behavior." Established --json cleanliness expectations across miner commands.

Distinct from:

• The CLI-JSON-output-routing cluster (Bug: noise in stdout interfere json parsing #956 → fix(cli): route warnings, errors, logs to stderr #957, fix: replace raw console.status with loading_context in CLI commands (#898) #1166, fix: route --json CLI commands through loading_context to suppress spinner #1150, fix(cli): route --json commands through loading_context for spinner consistency #1097, cli:Unify machine-readable CLI output #1169, etc.). Those all touch where logs go (stdout vs stderr). This issue is about a crash in the shutdown path triggered by valid advertised input. Different gate, different category.
• Defensive-guard closures [Bug] git issues list --id can crash on string/invalid bounty fields after shared _fill_percent() change #1078 / [Bug] resolve_network() crashes on non-string config network before ws_endpoint fallback #1084 / fix: reject negative and non-finite weights at load time #762 — those required hand-mocking invalid inputs into otherwise-typed fields. This crash uses the command's officially advertised --log-level warning flag with no mocking.

3-axis search (_drain_logs × TransitionNotAllowed × --log-level warning) returns no overlapping open issue or PR.