[cnicholas]

Summary

• What: bugfix in three signal rules, and corrected chance rates and wording on the WECO rules page.
• Why: an external review found the page's Rule 7 rate wrong (0.003% where it is 0.33%). Checking every rule found six of eight rates wrong, and three rules in the code that did not do what the docs, the rule descriptions, RULE_MIN_OBSERVATIONS, Nelson and Wheeler say: Rule 5 fired only on 7 points (6 rises), Rule 6 only on 15 alternating points, and Rule 2 ignored points beyond 3σ. The existing tests used 16-point sequences, so the boundaries were never pinned.
• Scope: signals/detectors.py (three rules), the Rule 2 description, boundary tests, validation/chance_signals.py (two more columns), weco-rules.md, the signal-detection notebook's prose, key-concepts.md, CHANGELOG. No new parameter.

Contract / Invariants (must remain true)

• Default behavior unchanged: chart, capability and loss numbers unchanged; Bishop gate 1547/0/4. Signal output on X charts changes by design (below)
• Residuals unaffected: no residual code touched
• Row/index alignment preserved: violation flags still mark the last point of each window
• Lane boundaries preserved: no plotting change
• Output schema compatible: SignalResult unchanged; the Rule 2 description text adds "or beyond"

Behavior Changes (explicit)

• New API: none
• New semantics:
  • Rule 5 fires on 6 consecutive rising (or falling) points; it needed 7
  • Rule 6 fires on 14 consecutive alternating points; it needed 15
  • Rule 2 counts any point more than 2σ out on the same side, including beyond 3σ
  • mR, Xbar and S use Rule 1 by default, so they are unaffected unless more rules are requested

Key Design Decisions

• Fix the code, not the docs: every written definition (docs, RULE_DESCRIPTIONS, RULE_MIN_OBSERVATIONS = 6 and 14, Nelson, Wheeler, Minitab test 5) agreed; the code was the outlier
• Exact rates: per-window probabilities for independent normal points, computed exactly (Euler zigzag numbers for Rule 6); per-chart rates measured by chance_signals.py, now with X under Rule 1 and Rules 1-4 so the rule-set table has a source
• Wording: a signal is a reason to look for a cause; "almost certainly", "process has shifted" and "is drifting" become "suggests"

Risks & Mitigations

• P0 risks: users comparing signal counts across versions see more X-chart signals
• Mitigations: CHANGELOG states the change and the size (70% to 79% of 100-point pure-noise X charts with a signal); rules= still selects any subset

Tests

Added / Updated

• TestRule5Trend: 6 rising points fire (once, at the 6th), 5 do not, 6 falling fire, a tie breaks the run
• TestRule6Oscillation: 14 alternating points fire (once), 13 do not
• TestRule2ZoneA: points at 2.5σ and 3.5σ fire, both sides, and with per-row limits

Regression Focus (must fail if broken)

• off-by-one in trend or oscillation length
• Rule 2 dropping points beyond the limit

Manual Verification

• pytest tests/: 2696 passed, 7 skipped
• python validation/e2e_bishop_report.py: 1547 passed, 0 failed; validation.md unchanged
• python validation/chance_signals.py rerun; weco-rules.md and the notebook table use its output
• pytest --nbmake docs/tutorials/signal-detection.ipynb; ruff check . clean; mypy shows the same 15 advisory errors
• Per-window rates: 0.27%, 0.31%, 0.55%, 0.78%, 0.28%, 0.46%, 0.33%, 0.010%

Notes

• Docs: weco-rules.md (rates, wording, measured rule-set table, an explicit target for the per-chart section), signal-detection notebook prose (cells 7, 22, 28), key-concepts.md Rule 2; CHANGELOG [Unreleased] Fixed
• Follow-ups (not in this PR): the notebooks' saved outputs, including signal-detection's, are refreshed in the notebook-refresh PR that follows this one