feat: backtest excursions by horizon and outcome by feature

[yanivil]

Description

Implements #92 on top of #103.

• Excursions by horizon per window, overall and per pattern: median MFE and MAE within 5 / 10 / 20 / 40 / 60 bars after the fill, in percent and in ATR, plus the share of signals that had reached the target, hit the stop or done neither by then. Medians, because a few runaway winners dominate the means.
• Features per replay row, measured at the scan day from the history the scan saw (row_features): close vs SMA200, SMA50 vs SMA200, the SMA200's change over 40 bars, distance from the SMA200 in ATR, stop and target distance in ATR, and bars from the pattern's last anchor (handle low, right shoulder, point 5, read from notes the way scan._drop_reason does) to the breakout. Nothing is added to scan.Signal, signals.json or the nightly report.
• Outcome by feature per window: traded signals bucketed on fixed edges (FEATURE_BUCKETS) by those features plus volume ratio, risk %, reward:risk and breakout age, each bucket with N, hit rate, mean and median R and the month-block interval; empty buckets omitted, rows without a value counted as missing. Fixed edges keep two replays, or the two windows of a --split, comparable bucket by bucket.
• summarise_rows is the former inner function of breakdown, now shared with the bucket table. JSON output gains features and horizons per window.
• Wiki 04 describes both tables; changelog.

Verified locally: ruff clean, 124 tests pass (4 new), and a 30-ticker --period 5y --days 250 --bars 500 run renders both tables and writes the JSON with the feature keys on every row.

Related Issues

Closes #92

Checklist

• Code follows existing project style and type annotations
• Threshold changes include a one-line "why" comment in scan.py (no thresholds changed)
• Offline unit tests pass (python -m pytest -q)
• Random-walk false-positive rate stays under 5% (detection logic untouched)
• Documentation updated in docs/wiki/
• CHANGELOG.md updated

🤖 Generated with Claude Code