Make the logistic regression demo runnable from a clone

[pH-7]

A fresh clone cannot start: the training notebook is empty and the API imports a model artifact that does not exist. Add a reproducible offline Iris training example, complete the notebook, pin the Python 3.12 dependencies, and generate the artifact inside the Docker build.

Define the four-feature API contract and return JSON 422 errors for malformed requests, including non-finite numbers that previously broke error serialization. The README describes the feature order, class labels, measured holdout evaluation, setup and demo limits. Iris is an explicit teaching-dataset choice for the previously empty training step; it is not a production-data claim.

Validation: nine tests pass, the notebook executes, actual localhost prediction/docs/error flows pass, Ruff formatting/lint and workflow lint pass, and the pinned runtime audit reports no known vulnerabilities. Training fits 120 rows and evaluates 30 held-out rows. CI now exercises the Python pipeline and Docker build/smoke check; Docker is unavailable locally. No model binary, secret, container registry publication or deployment is included.