Kevin's signals are mid-long term (weeks/months) and he uploads almost
exclusively pre-market or evenings. Before this change, every such
signal hit RiskManager.outside_market_hours, got consumed off the
Redis stream, and was lost. End result: 71 emitted signals, 0 trades.
New behaviour: when RiskManager rejects with outside_market_hours,
push the signal into a Redis sorted-set keyed by next_market_open
(via Alpaca's clock API — handles weekends + holidays). A background
drain task polls the set every kevin_defer_drain_interval_s (60s);
any signal whose target <= now gets re-run through process_signal.
Safety:
- kevin_max_defer_hours (default 72h) caps signal staleness so we
don't trade on week-old views.
- Other RiskManager rejections (cooldown, kill-switch, drawdown
halt) fall through to the existing drop path.
- kevin_defer_outside_market_hours toggle defaults True; flip to
false for legacy behaviour.
Slack: new notify_deferred() emits "🕒 Meet Kevin: DEFERRED
NVDA until Mon 13:30 UTC (market closed; conviction 0.85)" instead
of the noisy outside_market_hours rejection spam.
Tests: 5 queue + 4 integration = 9 new, all 32 trade-executor tests
GREEN.
