Double-bottom detector (experimental) and a --patterns switch for the replay

[yanivil]

Summary

The first candidate pattern since the cup went watch-only: a double-bottom detector, built on the inverse head and shoulders template, kept out of the nightly scan until it passes the replay gate (#126).

Detector detect_double_bottom

• Two consecutive swing lows within 3 % of each other, either one the lower (the shake-out is allowed); the rally peak strictly between them at least 10 % and 2 ATR above the higher low; the peak not too close to either low in time; a prior decline of one height into the first low or SMA50 below SMA200.
• A daily close above the peak is the trigger (constant, through evaluate_breakout, floor the second low); stop under the second low, the pattern's invalidation, like the H&S right shoulder; target one height (peak minus the lower low) above the entry; the H&S's age lag (the second low is a swing low, visible five bars later), volume rule, risk cap and listing limit.
• Bulkowski's 10 % rise rule is what keeps noise out: with the ATR depth rule alone 43 of 200 random walks show a W; with it, five (the inverse H&S: two). A test pins that rate.
• With equal lows the reward is one height and the risk slightly more, so the reward:risk sits just under 1 and tuned's floor keeps only the higher-second-low form; the replay measures both.

Plumbing

• PATTERN_DETECTORS holds every detector by name; ACTIVE_PATTERNS is what the nightly scan runs (unchanged: cup, H&S, Wolfe) and what the report and meta iterate. The double bottom is registered but not active.
• tools/backtest.py --patterns cup,ihs,wolfe,db (or the full names) replays any subset; the workflow gains the patterns input; the JSON records the patterns run.

Tests (164): levels recomputed from the notes and the documented formulas, single-rule mutations rejected (lows too far apart, rally under 10 %, lows too close in time, nothing to reverse), isolation (not in DETECTORS, the other fixtures are not W's and the W is not an H&S or a cup), the controls and the 200-random-walk rate, and --patterns parsing and selection end to end.

Docs: wiki 02 (the pattern's section, marked experimental), wiki 03 (its constants, the --patterns switch), wiki 04, README (the pattern table), changelog.

Replay: the eleven yearly point-in-time runs are dispatched from this branch with patterns=cup,ihs,wolfe,db, plus a second set with MIN_REWARD_RISK=None; the result and the decision go on #126. Merging this PR changes nothing in the nightly scan.

Test plan

• ruff clean, 164 tests pass locally
• tests workflow and the backtest smoke run on this PR
• the eleven-year replays (Double bottom: eleven-year replay of a candidate pattern #126)

🤖 Generated with Claude Code