12. Expose score-only inference mode (skip anomaly map computation) #17
Description
Description
Some use cases only require global or hybrid anomaly scores and do not need
pixel-level anomaly maps.
This issue proposes adding an optional inference mode that skips anomaly map
computation to reduce runtime and memory usage.
Scope
- Add a flag to inference call (e.g.
return_map=False) - Default behavior remains unchanged
- Scores must remain identical to full inference
Acceptance Criteria
- When disabled, anomaly map is not computed or returned
- S_F, S_P, and S_H remain numerically identical
- No breaking changes to existing API
Notes / Constraints
- Must not change model behavior
- Implementation must be explicit and readable
Testing Requirements
- Test equivalence of scores with and without map computation
- Test that anomaly map is omitted when disabled