Agentic shadow-trading board (paper only)

SCADS-hosted models given real tool-use trading agency -- unlike the /edge-board Kelly-sizing ledgers (a probability sized by a fixed formula after the fact), each model here decides for itself whether and how much to trade with genuine cash/position accounting, real risk guards, and the same bounded ReAct tool loop PredictionArena-style agentic boards use. Shadow (paper) trading only -- no real money is ever placed; see scripts/agent_trading_tick.py for the enforcement.

• leaderboard: one row per model -- account value, cash, realized/ unrealized P&L, return %, win rate over settled markets, open positions, trade count.
• equity_curves: per-model cash-only book-value curve (see agent_trading_stats.agent_equity_curve for exactly what this approximates and why) plus Sharpe/max-drawdown computed over it.
• recent_activity: a merged, newest-first transparency feed of trades, settlements, per-cycle theses, and notes across every model.
• eligibility: per model, whether its shadow track record currently clears a conservative, adjustable bar (settled trades, Sharpe, drawdown -- see agent_trading_stats.compute_promotion_eligibility). Purely observational: nothing reads this to grant any capability.
• forecast_learning: final-outcome Brier diagnostics, including up to four source-labelled weather cohorts per model. These are descriptive calibration signals, never sizing overrides.
• weather_operations: bounded recent counts from source-verified weather candidates through research, forecast, paper trade, and final resolution. It is operational coverage, not a performance claim.
• model_health: per model cycle heartbeat and account-ledger freshness. This separates a completed no-trade turn from a delayed, stale, or unverified model ledger.
• operational_health: bounded roll-up of provider degradation versus ledger-maintenance attention, without obscuring individual model status.