Without bootstrap, the brain learns purely from going-forward live ticks — meaning ~10 resolved trades per symbol per day. It takes WEEKS for BSS to converge, weeks for calibration to be honest, weeks for per-symbol blends to mean anything.

The bootstrap fixes this:

1. On first visit, pulls 60 days of daily OHLC bars from Stooq for every tracked symbol (SPY, QQQ, NVDA, BTC, etc.) — free, no key
2. For each (symbol, day) pair, reconstructs the same 22-feature vector the brain uses live: RSI, ATR%, vol ratio, distance from 50/200 MA, SPY change, setup flags from price action, etc.
3. Outcome label = next day's close direction (LONG win if next close > today's by 0.3%+, SHORT win if -0.3%-, dropped if within noise)
4. Feeds (features, outcome) to every learning module: ModelStore (Adam SGD), Calibrator (Platt), IsotonicCalibrator (PAV), RegimeCalibrator, BrierSkill, SharpeTracker, SymbolSkill, SymbolSharpe, SectorPerf, ReliabilityDiagram, PredictionHistogram, SymbolBias
5. Fits all calibrators on the populated pools so probabilities become honest immediately

Result: ~660 training examples (22 symbols × ~30 valid days/symbol after dropping flat outcomes) → BSS converges to a measurable value within minutes of opening the site. Per-symbol stats become meaningful. The brain is warm-started instead of cold.

Gated by localStorage flag so it only runs ONCE per browser. Reset the flag to re-run (e.g. if you clear training history and want to re-bootstrap). Each run takes ~10 seconds (24 symbols × 300ms delay).