Fix: consensus detection OOMs the analytics container on a username join

[mpont91]

ConsensusDetector._get_smart_money_trades joins the trades of one market to the scored traders on username:

FROM polybot.aware_global_trades t
INNER JOIN (
    SELECT username, total_score
    FROM polybot.aware_smart_money_scores FINAL
    WHERE total_score >= {min_total_score}
) s ON t.username = s.username

Practically no Polymarket wallet has a username set. On a live instance all 1,793 scored traders above the threshold have it empty, and so do almost all trades. The join therefore matches empty string to empty string and becomes a cartesian product.

Measured on the busiest market of the last 48 hours:

join on username        15,267,395 rows
join on proxy_address            the handful that are real

There is no LIMIT, and every row is materialised into a Python dict. The loop then repeats for ~100 markets.

Effect

The container is OOM-killed part way through consensus on every run. On the instance where I found this: RestartCount 56, and grep -c "AWARE Analytics - Complete" over the entire log returns 0 — no cycle has ever finished.

Consensus is step 7 of 12, so everything after it has never run: ML enrichment, edge persistence, drift monitoring, notification dispatch, and the strategy P&L job. Any dashboard fed by those tables shows whatever was last written by a manual run.

Fix

Join on proxy_address, which is always populated and is the key the rest of the scoring pipeline uses. _aggregate_positions is keyed on the address too — keyed on an empty username it collapsed every trader into a single bucket, so the trader count behind a signal was meaningless.

After the fix the same pipeline completes in 70.6s.