Harden SentinelML monitoring pipeline, API, and regression checks

[Ayushdevo]

Summary

• Harden PSI, feature schema, missingness drift, training input, SHAP output, and model-health handling.
• Make a fresh pipeline run create its missing baseline model; align API inputs to the trained feature order and add readiness/error responses.
• Add nine regression checks, CI, declared API/test dependencies, and corrected setup commands.

This branch contains 30 focused commits. The final Git tree was verified to match the locally tested checkout.

Verification

• .venv/bin/python -m src.run_pipeline completed with existing sample data/model artifacts.
• .venv/bin/python -m pytest -q: 9 passed.

The generated report artifacts were restored after verification and are not included in the change.

[Ayushdevo]

Post-merge code review: the hardening work improved startup safety, schema checks, PSI handling, and regression coverage. I found three operational gaps worth following up:

1. #2 — /ready validates artifact existence, not loadability/parsing, so corrupt artifacts can still report ready.
2. #3 — load_model() is process-cached indefinitely; replacing the model artifact can leave inference serving the stale model.
3. #4 — drift detection catches missing expected features but silently ignores unexpected production-only columns, which can hide schema drift.

These are follow-up improvements rather than reasons to revert the merged hardening PR.