enhancement: establish calibrated prediction-confidence and explanation boundaries

[CoreyLeath-code]

Audit finding

The repository provides a PyTorch/FastAPI seven-class facial-expression classifier and already states that expression does not reliably reveal internal emotional state. Searches found no committed reasoning or confidence_score implementation. Raw scores must not be labeled calibrated confidence or human-emotion certainty without a reproducible protocol.

Evidence inspected

• README.md: FER-style classification with a research/demo, non-medical boundary.

• Tests cover model, API, pipeline, and production-context paths.

• Existing issue Evaluation protocol audit #26 tracks evaluation protocol; this issue complements it.

Proposed enhancement

Separate predicted expression label, model score, calibrated confidence only when supported, model version, input-validation result, and an abstain/low-confidence state. Explanations must remain concise and must not claim mental state, intent, or pixel-level causality.

Acceptance criteria

• Preserve inference behavior and return calibrated confidence only after reproducible calibration.

• Test invalid images, low confidence, abstention, schema validation, and model-version metadata.

• Document calibration split, method, and limitations before publishing metrics.

• Keep optional context retrieval separate from image-model evidence.

NEEDS HUMAN DECISION

Approve the dataset, subgroup review, and calibration target before enabling a numeric confidence claim.