Keep submitting last-good payouts during reward database outages

[dylanuys]

Summary

• Propagate reward database read failures instead of treating them as an empty reward window.
• Persist the last-good payable scores by hotkey. On reward-data failure, leave scores unchanged and continue submitting weights from that snapshot, while refreshing discriminator kings normally.
• Resolve cached hotkeys to current UIDs so replacements cannot inherit rewards; exclude positive EMA history that was not eligible for payout.
• Keep genuine empty-window and intentional full-burn behavior unchanged. Skip only when no safe payout snapshot exists.

Fixes the September 27 incident where a SQLite read error caused zero generator weights to be submitted, without stopping validator weight updates during subsequent database outages. This does not repair underlying storage faults.

Existing snapshots remain compatible and keep their EMA history. The first successful scoring run populates the new fallback; no database migration is required.

Validation

• 242 tests passed; CI lint and diff checks passed.
• Covers database errors/partial reads, repeated fallback, unchanged scores, restart persistence, UID reassignment, excluded EMA history, fresh kings, legacy/invalid snapshots, legitimate empty payouts, and recovery.