feat(validator): pin miner rate + address to the reservation

[anderdc]

Swap 79 — what spawned this

Diagnosing allways swap 79 surfaced that a reservation pins the swap amounts
(to_amount/tao_amount/from_amount) but not the miner's rate or its
deposit/fulfillment addresses. At initiate (~20 min after reserve, gated by
BTC confirmation) the validator re-reads the miner's live commitment — so a
miner can move its commitment in the reserve→initiate window and the swap still
settles against the moved values.

The issue

Two exploits follow from handle_swap_confirm re-reading the live commitment
(allways/validator/axon_handlers.py — load_swap_commitment at the
commitment = ... call inside the axon_lock block):

• Rate-swing — a miner moves its rate between reserve and confirm; the
  selected_rate_str flowing into scale_encode_initiate_hash_input and
  vote_initiate is the moved rate, shortchanging the user.
• Address theft (total loss) — a miner changes its committed deposit address
  after the user has sent BTC. verify_transaction is then called with the new
  miner_from_address (axon_handlers.py, the provider.verify_transaction(...)
  call), vote_initiate fails verification, no swap is created, no slash fires —
  the miner keeps the BTC.

The fix

A validator-side, event-driven pin index — no contract change, consensus-safe:

• state_store.py — new ReservationPin dataclass + reservation_pins table
  (purely additive, no migration) with upsert/get/remove/purge/update
  methods and delete_hotkey cleanup.
• event_watcher.py — the watcher gains netuid/subtensor; a new
  MinerReserved branch reads read_miner_commitment(..., block=R) at the
  canonical reservation block R and persists the commitment snapshot.
  SwapInitiated/SwapTimedOut clear the pin; ReservationExtensionFinalized
  bumps its TTL.
• axon_handlers.py — handle_swap_confirm resolves rate + addresses from the
  pin (synthesizing a MinerPair and feeding the existing
  resolve_swap_direction), falling back to the live commitment when no pin
  exists.
• forward.py — purge_expired_reservation_pins() beside the existing
  pending-confirm purge.
• neurons/validator.py — wires netuid/subtensor into the watcher.

The index is keyed on the canonical block R and a canonical commitment read,
so every validator derives a byte-identical pin → an identical request_hash.
A validator with no pin falls back; it never produces a wrong pin. On any
failure (transient RPC, pruned block) no pin is written.

Review notes

• This is consensus-path code — it changes the provenance of the
  vote_initiate hash inputs.
• Needs a coordinated validator upgrade: mid-rollout, un-upgraded validators
  hash from the live commitment and upgraded ones from the pin; if a miner moved
  its rate/address the fleet can split on vote_initiate. The exploit closes
  only once a quorum runs the pin.
• Not yet testnet-tested — unit-tested only (state-store round-trip/purge,
  watcher MinerReserved/lifecycle, axon address-theft + rate-swing
  regressions). The plan's integration test (reserve, move the miner's
  commitment, confirm) still needs a testnet run.
• Pre-existing reservations made before this deploys have no pin and fall
  back to the live commitment — chosen over rejecting so in-flight swaps aren't
  broken. The exploit window stays open only for reservations straddling the
  deploy.