Baseline artifact pipeline phase 3: deploy consumes the store (image fetch, deployed markers, prewarm retired)

[anth-volk]

Phase 3 of the baseline artifact pipeline (phase 1: #643, phase 2: #648).

Phases 1–2 fill and verify a content-addressed GCS artifact store on every deploy; nothing consumes it yet. Phase 3 is the cutover:

• Replace the in-image dataset prebuild layer (prebuild_country_datasets) with a self-contained fetch_artifacts layer that downloads every manifest-listed artifact (3 datasets + 60 cohort baselines, ~430 MB) from the store into POLICYENGINE_DATA_FOLDER, with a version freshness gate and loud failure on any missing object.
• The deploy job consumes the precompute job's manifest_digest output: app.py resolves the manifest from GCS on the runner and passes its content as the fetch layer's args, so the Modal layer cache busts exactly when the artifact set changes.
• Each deploy leg writes a deployed/<environment>.json marker after the health check — the liveness signal for the phase-4 GC.
• Retire the manual prewarm ritual: delete prewarm_app.py, the prebuild layer, and the force_build escape hatch; migrate the coupling tests.

The merge is atomic by design: the producer already flows from phase 2, so this PR only adds the consumer.

Phase gate (post-merge): deploy green on both legs with the image build fetching in seconds instead of building; a beta economy request logs baseline_artifact=hit with p50 down; markers visible in the bucket for beta and prod; a fault drill proving delete-then-rerun healing.