edge-margin policy + entry_price_max cap (hybrid: 2 Optuna params)

edge-margin policy: single edge_threshold (model_pred - market_ask) instead of 3 hard gates

btc: cap back to 0.60, n_trials 150 to 300 for variance reduction

btc only: bump entry_price_max cap to $0.70

raise entry_price_max cap to $0.60 (more coverage, trades can enter pricier tokens)

fix KeyError: directional-mode diagnostics (no more entered/hedged/exit_type columns)

directional classifier: predict UP wins + price-aware trading (lower ask = better EV)

label v4 baseline run: post-hoc optimal hedge ceiling, no policy Optuna

add policy-parameter Optuna: tune (entry_thr, combined_thr, min_tick) before classifier

v1_arb: simulate real live policy (entry ≤$0.40, hedge at combined ≤$0.80)

label v4: entry FIXED at tick 120, hedge scan [121,899], threshold=$20

label: entries scan [120,899] (live reality) + threshold=$70 (75%ile)

label: min_entry_tick=120 (no leakage) + threshold=$10 + pnl distribution log

label fix: binary uses hedged-arb only (no-hedge=directional bet, not arb)

clean import in iter_orderbook_slug_pairs

stream orderbook row-groups via pyarrow (fix OOM, peak 5MB/slug)

fix: numpy 2.4 float(1d-array) in _walked_ask_cost

log full traceback on window error (debug float/int conversion)

batch-load orderbook (fix OOM on 37GB BTC parquet)

v1_arb trainer code for btc

initial commit