fix(validator): non-earner trace reads live quotes, not dead last_known_rates

[anderdc]

Problem

The scoring trace labelled every non-earner `reason="no_rate_posted"`, regardless of actual cause. `non_earner_lines` read `getattr(self, 'last_known_rates', {})` — but every writer of that attribute was deleted in the B3.6 substrate teardown (87a6732). The `getattr` default masked the dead reference, so the dict was always empty and `diagnose_non_earner` short-circuited on its first branch.

Seen on testnet 2026-08-22: uid 12 (50 ok / 3 fail → struck out, correct per #699) with a live sol→btc quote logged as `no_rate_posted eligible=False`, which sent a debugging session chasing a restart/rate-replay theory that didn't exist. The true reason was `ineligible`.

Fix

• `calculate_miner_rewards` already fetches `live_quote_rates(...)` for the quote reconcile; hoist that result and pass it as `live_rates` through `log_scoring_trace` → `non_earner_lines`.
• `non_earner_lines` builds its per-hotkey rate map from those 4-tuple lanes, collapsing the backing dimension (the diagnosis already anchors on the pair's hub leg via `hub_leg()`).
• Dead `last_known_rates` lookup removed.

No change to reward math or crown logic — diagnostic text only.

Tests

`TestNonEarnerLinesUseLiveRates`: a struck-out miner with a live quote reads `ineligible`; a miner with no live quote still reads `no_rate_posted`.