yzhao062 / pyod / 25592119807

94%

master: 94%

Build Type

push

github

Committed by web-flow

Commit Message

refactor: ADEngine type hints, validation, exception handling, NL parser (#670)

* refactor: ADEngine type hints, validation, exception handling, NL parser

Tightens the post-decomposition codebase. Implements PR 3 of the
ADEngine refactor sequence (PR 1 = stability fix, PR 2 = decomposition,
PR 3 = tech debt cleanup).

Behavior tightenings (3 user-visible changes documented in CHANGES.txt):
- iterate(feedback) raises ValueError on malformed feedback dicts
  (previously: silent confirm_with_user fallthrough)
- Per-detector exceptions in run() and analyze() emit WARNING logs
  before swallowing
- NL feedback parser uses word-boundary regex (re.search with \b)
  instead of substring `in`; some incidental matches no longer fire

Internal cleanup (no behavior change):
- Type hints across all 5 modules + ad_engine.py public methods
  (no mypy enforcement gate; documentation only)
- Magic numbers extracted to named constants:
  _VERDICT_HIGH_THRESHOLD / _VERDICT_MEDIUM_THRESHOLD /
  _SINGLE_DETECTOR_AGREEMENT_FALLBACK in _quality_metrics.py;
  _CONTAMINATION_INCREASE_FACTOR / _CONTAMINATION_DECREASE_FACTOR /
  _CONTAMINATION_MAX / _CONTAMINATION_MIN /
  _NL_HIGH_CONFIDENCE_THRESHOLD in _nl_feedback.py
- QualityDict TypedDict in _quality_metrics.py
- validate_structured_feedback() raises on missing/unknown action,
  missing required fields, wrong type
- compute_feature_importance / feature_contributions narrow their
  bare except Exception to (AttributeError, ValueError, TypeError)
  with logger.debug
- NL parser refactored to declarative pattern table + parse helper
- Contamination math deduplicated to adjust_contamination_up/down

Tests: 13 new tests across 3 classes (TestValidateStructuredFeedback 5,
TestNLPatternTable 7, TestExceptionLogging 1). 135/135 pass.

4 parallel subagents extracted helper-module changes; orchestrator
sequentially handled ad_engine.py + test additions. Plan-reviewed
and implement-reviewed by Codex over 2 rounds.

Spec: docs/super... (continued)

Coverage Stats

217 of 224 new or added lines in 6 files covered. (96.88%)

4 existing lines in 3 files now uncovered.

16304 of 17426 relevant lines covered (93.56%)

10.29 hits per line