Fix neuron lifecycle: instance-level state, scores lock, autoupdater guards

[dylanuys]

Summary

Three small correctness fixes in the neuron lifecycle, all found by inspection. No behavior change in the normal single-instance path.

• BaseNeuron shared state — exit_context, block_callbacks, and the substrate handles were declared as class attributes, so they were shared across instances: a second instantiation reused the same exit context and appended to the same callback list. They're now initialized per-instance in __init__. Also switched the not-registered path from the builtin exit() to sys.exit(1), since exit() isn't guaranteed present outside the REPL.

• update_scores lock — the score-array extension, EMA update, and inactivity zeroing now run under the existing _state_lock. Previously they mutated self.scores without the lock while save_state read it under the lock.

• Autoupdater — locate the repo root by walking up to the first .git directory instead of looping until a directory literally named bitmind-subnet (which never terminates when the repo is cloned under a different name). The git pull now runs via subprocess with --ff-only, and a non-zero exit aborts the update instead of proceeding as if it succeeded.

Test plan

• pytest tests/ — passes
• ruff (error-level) — clean
• Testnet validator restart + autoupdate cycle before the next release

🤖 Generated with Claude Code