[arch] image-service: production ML pipeline with model hot-swap

[cofade]

Background

docs/01-introduction-and-goals/README.md notes that the image-service classifier is a stub today and that real ML integration is "planned." This issue makes that pipeline real and operable so that downstream features can plug in without a Dockerfile rebuild for every model swap.

Why it matters

Almost every feature on the wishlist (cell-occupation timeline, hatching prediction, species ID, time-lapse triage) depends on a real model running in image-service. Without a hot-swap-capable pipeline, every new model is a redeploy.

Proposal

1. Model directory mounted as a volume: /models/<name>/<version>/{weights, config.json, README.md}.
2. Model registry loaded at startup, with optional reload-on-SIGHUP (or admin endpoint) for swap-in.
3. Inference contract: each model exposes a common predict(image) -> {detections, scores, model_version} interface.
4. Telemetry: every inference logs model_version so downstream consumers can filter or retroactively understand drift.
5. First real model: a small object detector (YOLO-nano-class or similar) replacing the stub — even baseline accuracy is fine; the point is the integration, not SOTA numbers.

Acceptance criteria

• /models/... volume mount in docker-compose.yml
• Model registry + at least 2 loadable models (the existing stub plus one new)
• model_version recorded with every output that crosses a service boundary
• New ADR in docs/09-architecture-decisions/ covering the choice + model lifecycle
• arc42 chapter 05 (image-service file) updated to describe the new internal structure
• Unit + smoke tests covering "swap a model without restarting the container"

Out of scope

• Choosing the production model — that is a follow-up tracked separately.
• GPU inference — CPU-only is fine for now (ESP-side latencies dwarf inference time).