Close parser-poison free-rider hole + harden sentinel-rate handling

[LandynDev]

Summary

Primary: closes a new free-rider exploit not covered by v1.0.6.

A miner posts a sane rate (e.g. 340), builds credibility, then overwrites their commitment with "x" (or any parser-rejected string: wrong version, NaN, malformed, unsupported chain). The Bittensor Commitments pallet has no delete extrinsic, but set_commitment is overwrite-only — a single follow-up tx puts the miner in the "vanished from poll" state. Today refresh_miner_rates only emits a 0-terminator inside its per-direction loop over admitted pairs, so a vanished pair leaves the prior positive rate as the latest in state_store. Scoring reads the ledger, finds 340, calls is_executable_rate(340, ...) → True (340 is a perfectly executable rate), credits the miner crown forever. They can never be reserved (live commitment read fails), so they free-ride emissions until deregistration.

v1.0.6's executability filter (PR #395/#420) does not catch this. That filter rejects rates whose value is unexecutable. The free-rider's stored rate is a sane, executable value — the bug is that the rate is stale and nobody emits a terminator when the commitment disappears.

Secondary: hardens sentinel-rate handling.

PR #395/#420 zeroed sentinel-rate posters' crown reward at the scoring layer. They still appear in alw view rates, get axon-pinged on reserves, burn RPC budget on every commitment poll, and have no defense-in-depth at the axon handlers. This PR drops them at the parser layer so they're invisible to the validator entirely.

Both exploits present identically to the validator — "admitted previously, no admitted pair this poll, stale positive rate in state_store" — so one sweep covers both.

What changes

• parse_commitment_data / read_miner_commitment / read_miner_commitments gain optional min_swap_rao / max_swap_rao kwargs. Defaults are 0/0 (permissive, matching is_executable_rate). Validator passes cached bounds; CLI callers stay unchanged.
• refresh_miner_rates gains a second sweep after its per-direction loop: any (hotkey, from, to) previously > 0 in last_known_rates but missing from this poll's admitted set gets a 0-terminator emitted. This is the free-rider fix — crown attribution stops at the next block whether the miner's commitment was parser-rejected for sentinel reasons or because they overwrote it with garbage.
• bootstrap_miner_rates hydrates last_known_rates from persisted state before seeding admitted pairs so the same defense fires on the first poll after a restart that crossed a miner's parser-poison flip.
• handle_swap_reserve and handle_miner_activate get defense-in-depth executability checks. handle_swap_confirm is intentionally left alone (the reservation-pin path already handles bounds-shift correctly).
• alw view rates drops unexecutable rows with a footer count. alw view miners keeps them but decorates with ✗. alw swap quote adds the "every miner is sentinel" summary branch. alw swap now pre-filters before showing the ranked table.
• New read_unexecutable_commitments helper returns hotkeys whose commitment parses permissively but drops with bounds. Not wired into any caller in this PR — staged for the follow-up auto-deactivation PR.

Not in this PR

Auto-deactivation via vote_deactivate is consensus-sensitive and lands separately. Until then, a parser-poisoned or sentinel miner stays active on-chain — they just stop earning crown, stop being reservable, and stop appearing in alw view rates. Operators can vote_deactivate manually if needed.

Test plan

• pytest tests/ — all pass (634)
• ruff format && ruff check clean
• Staging validator with a planted sentinel commitment: confirm view rates hides them, view miners shows ✗, swap quote labels unexecutable, handle_swap_reserve rejects.
• Staging validator with a planted parser-poison sequence (post sane rate, accrue events, post "x"): confirm next poll emits 0-terminator and scoring stops crediting at the terminator block.
• No cross-validator divergence over a full scoring round.