fix(validator): serialize TAO source-balance check under axon_lock

[anderdc]

Root cause

The validator logs recurring cannot call recv while another thread is already running recv or recv_streaming errors. Root cause: a single substrate websocket call was made outside the lock that serializes that connection.

In handle_swap_reserve (allways/validator/axon_handlers.py), the source-balance check provider.get_balance(synapse.from_address) ran outside the with validator.axon_lock: block. The comment claimed the source-chain RPC is "a separate connection from substrate, so it doesn't need axon_lock."

That is true for a BTC source (Esplora/Maestro = HTTP) but false for a TAO source: the subtensor provider's get_balance calls self.subtensor.get_balance(...) on axon_subtensor — the exact websocket axon_lock exists to serialize. So every TAO->BTC reserve raced the lock-protected readers (other axon handler threads + the forward loop's bounds_cache reads on axon_subtensor), tripping the recv collision.

Fix

• Add uses_substrate: bool = False on the chain-provider base (base.py); set uses_substrate = True on the subtensor/TAO provider (subtensor.py). BTC stays at the default False.
• Gate the balance check on the flag: serialize the TAO read under axon_lock, keep BTC's HTTP read lock-free so a ~300ms Esplora call can't stall the forward loop. (axon_lock is an RLock, so the later re-acquire in the main block is fine.)

if provider.uses_substrate:
    with validator.axon_lock:
        balance = provider.get_balance(synapse.from_address)
else:
    balance = provider.get_balance(synapse.from_address)

Other handlers scanned

Scanned handle_miner_activate and handle_swap_confirm for the same pattern (any provider.* / axon_subtensor.* substrate call outside axon_lock). No additional leaks found — both wrap all their substrate work (is_hotkey_registered, read_miner_commitment, contract reads/votes, get_current_block, verify_transaction) inside with validator.axon_lock. The only leak was the reserve-time balance check.

Tests

Added TestSourceBalanceLock in tests/test_axon_handlers.py:

• asserts SubtensorProvider.uses_substrate is True, BitcoinProvider.uses_substrate is False, base default False;
• handler-level test that a TAO-sourced reserve holds axon_lock around get_balance, and that a BTC-sourced reserve does not.

Full suite: 689 passed. ruff check + ruff format --check clean.