feat: replay the 1H Wolfe Wave specification on its own terms

[yanivil]

Context / why

An external 1H Bullish Wolfe Wave specification was put next to scan.detect_bullish_wolfe. The two share the geometry and differ on almost everything around it: hourly bars instead of daily, swing points three bars deep instead of five, a fixed 25-session window, a stop half an ATR under point 5 instead of a quarter, the point-4 high as the only target instead of the EPA, a binary exit with a time stop instead of the listing rules, and no reward:risk floor.

None of the scanner's replay evidence transfers to that trade, and scan.py's pipeline (last-bar alignment, the pivot lag, the evaluator) assumes daily bars, so bending it to the specification would test neither. This adds a separate harness that implements the specification as written and logs the seven factors of the accompanying review, so each can be judged by its buckets the way tools/backtest.py judges features.

Summary of changes

tools/backtest_wolfe_spec.py (new): the replay, with its own constants block.

• Structures: strict three-bar swing points, consecutive swing lows 1-3-5, the highest swing highs between them as 2 and 4, the price hierarchy High(P2) > High(P4) > Low(P1) > Low(P3) > Low(P5), both lines falling with 2-4 the steeper.
• Trade: entry at the first close back above line 1-3 after point 5 swept it, stop at the point-5 low minus 0.5 ATR(14), target the point-4 high, exit at whichever comes first (a bar touching both is a stop, gaps fill at the open), else a time exit; one position per symbol; no reward:risk filter.
• Per trade it records the specification's logging schema plus the review's seven factors: trigger volume over the 20-bar average, RSI 14 at point 5 against point 3, the trigger bar's body over its range, the projected reward:risk, the close against a higher-timeframe EMA 50, an earlier swing low within 0.5 ATR of point 5, and SPY against its SMA 50 with the VIX.
• The report gives the funnel from structures to trades, the headline win rate (TP1 hits over closed trades), the outcome per factor bucket and per year, and states the rules it ran.
• Where the specification is silent it is read conservatively: the entry is never anticipated (the reclaim counts only from the bar that confirms point 5, with the literal lag kept for comparison), the reclaim must come within 25 bars, the structure must fit the window up to the entry bar, and a lower low before the reclaim makes that low the structure's point 5. The specification's rolling-max pivot sketch would also mark every bar of a flat stretch, so ties are excluded as its prose requires.
• Runs on any yfinance interval. Yahoo serves intraday bars for the last 730 days only, so a longer --years on 1h is clamped and the report says so.

test_backtest_wolfe_spec.py (new): 19 offline tests on deterministic hourly fixtures, covering the swing rule, each geometry invariant one violation at a time, a fixture yielding exactly one trade with the specification's levels, the non-anticipated entry, every trigger-rejection reason, the exit rules including gaps and the time exit, low-R trades kept and bucketed, one position per symbol, a random-walk ceiling, RSI and ATR equal to scan's, the bucket tables, the report and JSON schema, the intraday clamp and main() through the yfinance stand-in.

.github/workflows/backtest-wolfe-spec.yml (new): runs the tool per interval (default 1h,1d, three years), tees the reports into the job summary and keeps the rows and reports as an artifact.

Docs: README layout and limitations, a [Unreleased] entry in CHANGELOG.md, a dated section on the tuning wiki page with the rules as implemented and the readings above, and rows on the testing page for the new test file and workflow.

No change to the scanner, its rules, the nightly job or any existing output.

Test steps

ruff check --select E,F,W --line-length 120 .
python -m pytest -q
python -m pytest test_backtest_wolfe_spec.py -q -s -k random_walks   # prints the noise rate

Both pass locally on the pinned Python 3.12 environment: lint clean, 183 tests (19 new), and the random-walk rate is 2.27 trades per 1,000 hourly bars against a ceiling of 4.

For real data, this branch's push starts the backtest-wolfe-spec workflow; the job summary then carries the funnel, the win rate and the factor tables for both intervals.

Doc impact

README.md, CHANGELOG.md, docs/wiki/03-Configuration-and-Tuning.md and docs/wiki/04-Testing-and-Contributing.md are updated in the same commit. The tuning page's new section is written to receive the measured results once a run exists; it currently states the rules and the readings only, and claims no figures.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01H6cvDsVigeuGGxgEMF7zEy

---

Generated by Claude Code

[yanivil]

github-advanced-security is red, and it is not this PR's failure. Standing down on it; here is the evidence.

The job never analysed the diff. It failed while creating its own agent session:

Error creating PR review request: SessionModelError: Execution failed:
CAPIError: 400 The requested model is not supported.
COPILOT_AGENT_MODEL: sweagent-capi:claude-opus-5[ReasoningEffort=medium]

That is GitHub's own code-scanning agent asking its API for a model the API rejects. It is not reachable from repository content, and the run ends before any file is read.

It is also not new. The check has failed identically on the last eight pull requests in this repository, over nine days, including every one that merged into main:

Run	PR	Conclusion
35362266175	#130 (this one)	failure
35265663195	#128	failure
35260821453	#127	failure
35258088021	#125	failure
35257143981	#124	failure
35236119696	#123	failure
34977667092	#122	failure
34330445863	#120	failure

No fix exists to port: the failure is in a GitHub-hosted agent's model selection, outside this repository's code and workflows. I have not re-run it, because eight identical failures across unrelated diffs are stronger evidence than a ninth would be, and a re-run spends Actions minutes on a known-broken third-party agent. If you want the check green, it needs raising with GitHub Advanced Security support rather than a change here.

Everything this PR does own is green on the current head: tests passes (lint, 183 tests, coverage), and backtest-wolfe-spec completed successfully, with its reports in the job summary and the rows kept as an artifact.

---

Generated by Claude Code