fix(signals): a short series no longer returns an unqualified all-clear (#114)

[cnicholas]

Summary

• What: SignalConfig.min_observations becomes a real advisory threshold, and SignalResult.summary reports partial evaluation instead of printing ✓ No signals detected on a series it could barely examine.
• Why: on a four-point X chart six of the eight rules cannot fire, and the one that can is weakened by any trend in the data. The summary still printed the checkmark, and min_observations was never read by the detector. Reported in Series length (T) as an axis in design-state admissibility — and a saturation question #114 and privately as GHSA-hw63-2x95-fmpv.
• Scope: signals/detector.py, signals/result.py, two docstrings, docs/reference/weco-rules.md, docs/reference/api.md, CHANGELOG.md, SECURITY.md version table, tests.

Contract / invariants

• Defaults unchanged — which rules run, and what they flag, is identical. The per-rule table is now SignalDetector.RULE_MIN_OBSERVATIONS and still gates rule execution. validation/e2e_bishop_report.py exits 0 with 280 of 280.
• Residuals unaffected — signals only.
• Row/index alignment preserved — no data path touched.
• Output schema compatible — to_json still writes a records array. SignalResult constructor additions are keyword-with-default; 4-arg positional construction still works and reads as complete.
• Pinned error strings unchanged — rules_skipped reason text is untouched. The clean-series ✓ No signals detected in X line is unchanged for a complete evaluation.

Behaviour changes

• detect_signals() on fewer than min_observations rows emits one ProcessBehaviorWarning per chart naming both numbers. Silence with warnings.simplefilter('ignore', ProcessBehaviorWarning).
• is_partial / evaluation_status are also partial when the series is below min_observations, even if every runnable rule ran.
• A partial evaluation with no violations prints ⚠ Partial evaluation in X: 2 of 8 rules applicable at n=4 (below min_observations=20); skipped: rule_3 (needs 5), ... No signals from the rules evaluated. With violations present, the banner summary gains an Evaluation: partial (...) line.
• New SignalResult attributes: rules_evaluated, rules_applicable, n_observations, min_observations, below_min_observations, evaluation_note. repr gains evaluation='partial'. The Excel Summary sheet gains Evaluation Status and Rules Evaluated / Applicable rows.

Methodology

• No methodology change — reporting and ergonomics only. Series length as a design-state concern is deliberately not in this PR; that is the second half of Series length (T) as an axis in design-state admissibility — and a saturation question #114 and waits on a methodology discussion.

Tests

• test_signals::test_insufficient_data_partial_evaluation — extended: carries the 3 of 8 denominator and the summary is not an all-clear.

• test_signals::test_below_min_observations_is_partial_and_warns — every runnable rule ran, series under threshold: warns naming both numbers, result partial.

• test_signals::test_at_min_observations_is_complete_and_silent — adequate series: no warning, checkmark summary unchanged.

• test_signals::test_min_observations_does_not_change_which_rules_run — threshold 1 vs 1000 give identical rules and violations.

• test_signals::test_signals_present_and_partial_shows_evaluation_line.

• test_signal_result_surface::test_summary_partial_is_not_an_all_clear, ::test_positional_construction_still_complete.

• test_documented_rule_coverage::test_docs_minimum_observations_table_matches_detector — the docs table is pinned to RULE_MIN_OBSERVATIONS.

• test_documented_rule_coverage::test_four_point_ramp_is_not_an_all_clear — the advisory's reproducer, end to end through pb.formulate.

• pytest tests/ — 2335 passed, 10 skipped

• ruff check . — clean

• mypy processbehavior/signals — clean

• Golden masters untouched

Notes

• RuleSet.custom(min_observations=) is still dropped by to_config(). Pre-existing, separate, not addressed here.
• run_rules_viz.py early-returns on has_signals, so a partial result draws no annotation. Not addressed here.
• The warning does not collapse across repeated calls under the default filter, because pandas resets the warning registry internally. One warning per chart per call is the honest description.
• Closes the advisory once it is published. Second half of Series length (T) as an axis in design-state admissibility — and a saturation question #114 (series-length advisory beside the design state) follows in a separate PR after the methodology discussion.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BKphCi1LC1hjGYteWEK9JV