Fix collateral-baseline zeroing that recycled emissions (v1.0.9)

[anderdc]

Problem

Subnet 7 emissions were recycling ~90–100% to RECYCLE_UID. Root cause confirmed on the live validator (sqlite + logs): the in-process event watcher records a miner's collateral as 0 whenever it applies a fee/slash delta with no prior baseline.

On SwapCompleted/SwapTimedOut, _apply_collateral_delta(block, miner, -fee/-slash) computes prior + delta, where _latest_collateral returned 0 for a miner with no collateral event, and _record_collateral_event clips negatives to 0. Smoking gun: UID 65's only collateral row was 0 @ block 8311293 — the exact block of its SwapCompleted #2073 — while axon-preflight logs showed ~0.48 TAO on-chain throughout.

The #409/#423 capacity / can_fund crown gate then reads that 0 (capacity_factor → 0, can_fund → false) and drops the miner from crown entirely → 0 emission. 8 of 14 tracked miners had latest collateral = 0, so crown only reached the few correctly-tracked miners and otherwise recycled. Scoring traces: window=[8312102,8312402] recycled=1.0, [8312402,8312702] recycled=0.90.

Secondary: the scoring-trace diagnostic mislabeled these miners as outbid (diagnose_non_earner never checked collateral and wasn't direction-aware — it printed UID 65's better lower-wins rate 279.3 vs best=280 as a loss).

Changes

• Source fix (event_watcher.py): _apply_collateral_delta skips when there is no known baseline (_latest_collateral now returns Optional[int], None = unknown) instead of fabricating a 0.
• Reconcile (event_watcher.py, wired in forward.py + neurons/validator.py): reconcile_collateral_from_contract resyncs active miners' collateral to on-chain truth at startup (heals the warm-restart path, which does no contract reads → fixes the corrupted rows on redeploy) and once per scoring round (corrects drift). Writes at the current block only, preserving the #409 per-block / no-post-window-top-up property.
• Fail-open gate (scoring.py): can_fund and capacity distinguish absent collateral (unknown → fail open: pass / capacity 1.0) from present-0 (genuinely withdrawn → excluded), in both the replay and the live snapshot. Safe because the contract auto-deactivates anyone below min_collateral, so an active miner always holds ≥0.1 TAO.
• Truthful diagnostics (scoring_trace.py): diagnose_non_earner is direction-aware (tao→btc lower-wins) and collateral-aware → reports insufficient_collateral / unknown_collateral / competitive_but_unfilled instead of a blanket outbid.
• Version: 1.0.8 → 1.0.9.

Tests

pytest tests/ — 657 passed. New/updated coverage:

• no-baseline fee must not fabricate a 0 row;
• reconcile heal / skip-inactive / idempotent / RPC-failure;
• fail-open on unknown collateral vs exclusion on known-zero (existing test_zero_collateral_zeros_reward updated to seed a real present-0 event, since the old collaterals={'hk_a':0} setup produced an absent series that now correctly fails open);
• the four diagnose_non_earner reasons.

Verification after deploy

On the validator: confirm the 8 zeroed hotkeys (incl. 5Hb1Lxqm…) repopulate in state.db, V1 scoring recycled drops below 0.90, active best-rate miners' per-uid line shows cap>0/nonzero reward, and leaderboard crownShare rises over subsequent rounds.