Add reproducible evaluation pipeline and stronger baselines

[mkarson1997]

Goal

Upgrade the project from a course-style notebook into a reproducible ML case study.

Completed portfolio/reproducibility foundation

• Added MODEL_CARD.md with intended use, out-of-scope clinical use and limitations.
• Added EXPERIMENT_PROTOCOL.md covering data provenance, deterministic splits, seeds, preprocessing, training configuration, required metrics and baseline comparisons.
• Added explicit MIT licensing for project-authored material.
• Added THIRD_PARTY_NOTICES.md separating dataset/framework rights from project licensing.
• Added PR checklist for reproducibility/responsible-use review.
• Added dependency-free scripts/validate_project.py.
• Added GitHub Actions ML Portfolio Quality workflow.
• CI validates that the notebook is readable/structured and that required model/reproducibility/security/license artifacts exist.
• Latest quality workflow is green.
• README now exposes the quality badge and local validation command.

Remaining experimental work

• Replace the tiny 16-image validation design with a stronger deterministic split.
• Record fixed seeds and exact environment/dependency versions for the next run.
• Report precision, recall/sensitivity, F1 and ROC-AUC alongside accuracy.
• Add a simple baseline for comparison.
• Add at least one transfer-learning baseline such as MobileNetV2 or DenseNet.
• Export/record metrics consistently so runs can be compared.
• Add confidence intervals and/or calibration analysis where practical.

Acceptance criteria

The issue should close only after a new reproducible experiment follows EXPERIMENT_PROTOCOL.md and the README presents an honest baseline-vs-model comparison without implying clinical readiness.