feat: backtest features from the second review, and a 1.25-ATR stop variant

[yanivil]

Description

The testable remainder of the second external review (2026-09-08), recorded as replay features under the protocol of docs/wiki/03, then tested over the same eleven years. No rule changes.

• Six features on every replay row and in the outcome-by-feature table: the pattern's depth in ATR (cup bottom to right rim, shallower shoulder to head, Wolfe point 1 to point 5); the breakout bar's close within its own range; whether that close cleared the prior bar's high; the breakout volume z-score against the prior 20 bars; and for cups the handle's mean volume over the cup's and the slope of the handle's volume as a fraction of its mean per bar.
• Grid: an extra 1.0 ATR row, 1.25 ATR under the structural low, so the review's 1.2-ATR stop is scored on the same signals as the others.
• Wiki 03, new subsection "A second review, tested": what the review proposed, what was already in place, and the ten-year result of every testable item. Changelog.

Results (eleven yearly point-in-time runs from this branch, 34255129970 to 34255999955, the second set with CUP_MIN_ROUNDNESS=0.8; the stop variant on the two-year grid run 34256064869):

Proposal	Ten-year result	Verdict
Volume z-score of 1.5 as a gate	monotone, +0.16 R at or below 0 to +0.34 above 3; the gate keeps 315 of 1630 signals at +0.32 and drops 1315 at +0.20	feature, not gate
Close above the prior bar's high	with +0.19 R on 1405; without +0.42 R on 225, positive in 9 of 11 years	reversed
Strong close within the bar	+0.24, +0.24, +0.21	no effect
Handle volume decay (cups)	+0.06, +0.01, +0.06 by ratio; falling slope +0.05, rising +0.01	no effect
Cup R² 0.80 instead of 0.70	cups −0.10 R on 127 against +0.03 on 307; the 180 removed ran +0.12	worse
Pattern depth in ATR	at or below 3 ATR +0.26 R with a 42 % hit rate; above 10 ATR +0.26 with 15 %	no gate
Stop 1.25 ATR under the low	+0.20 R at 49 % against +0.29 at 41 %; same order in both windows	worse per trade

Verified: ruff clean, 135 tests pass (new arithmetic checks for every feature on the cup fixture, per-pattern depth for inverse H&S and Wolfe, cup-only features absent elsewhere, 24 grid rows and the new row rendered); a 30-ticker real-data run renders the buckets.

Related Issues

Related: #109 (cups: the roundness and handle-volume results bear on it), #108 (drawdowns), #96 (protocol)

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