feat: add Docker support for one-command local setup

[adarsh-7-satyam]

Closes #45

Problem

Setting up this project locally requires contributors to manually install Python 3.8+, Node.js 18+, create a virtual environment, and run bash.sh — which doesn't work on Windows without WSL. Version mismatches cause hard-to-reproduce environment-specific bugs, creating friction especially during high-volume programs like DMP 2026.

Solution

This PR adds full Docker Compose support so any contributor can get the
entire stack running with:

cp backend/.env.example backend/.env
# fill in GITHUB_TOKEN and OSDG_TOKEN
docker compose up --build

Frontend → http://localhost:3000
Backend → http://localhost:5000

---

Files Added

File	Purpose
docker-compose.yml	Orchestrates backend + frontend services
backend/Dockerfile	Python 3.11-slim image, installs requirements.txt
frontend/Dockerfile	Node 18 Alpine image, npm ci
backend/.env.example	Documents GITHUB_TOKEN and OSDG_TOKEN for contributors
.dockerignore	Excludes venv, node_modules, .env, .next from images

Code Changes

File	Change
backend/app.py	Added GET /health route required by Docker healthcheck
frontend/services/api.ts	Backend URL now reads NEXT_PUBLIC_API_URL env var with fallback to 127.0.0.1:5000 so bare-metal workflow is unchanged

---

Design Decisions

HuggingFace model cache — facebook/bart-large-mnli and sentence-transformers/all-MiniLM-L6-v2 are large models that download at runtime. A named volume hf-model-cache persists them across container restarts so they download only once.

Flask bound to 0.0.0.0 — Flask defaults to 127.0.0.1 which is unreachable from outside the container. The backend Dockerfile uses flask run --host=0.0.0.0 to make it accessible on the Docker network.

Volume mounts for live reload — Source directories are mounted so code edits reflect immediately without rebuilding the image. Anonymous volumes guard node_modules and __pycache__ from being shadowed by the host.

depends_on with healthcheck — The frontend waits for the backend /health route to return 200 before starting, preventing cold-start race conditions.

Non-breaking fallback in api.ts — The ?? operator means contributors not using Docker get the same 127.0.0.1:5000 behaviour as before. Zero regression for existing workflow.

---

Testing Done

Tested locally on Windows 11 with Docker Desktop 27.4.0:

Level 1 — file verification

• All 5 new files confirmed present on disk
• /health route confirmed in backend/app.py line 17
• NEXT_PUBLIC_API_URL confirmed in frontend/services/api.ts line 8

Level 2 — image builds

• docker build -t unsdg-backend ./backend — 6/6 steps FINISHED
• docker build -t unsdg-frontend ./frontend — 12/12 steps FINISHED

Level 3 — full stack smoke test

• curl http://localhost:5000/health → {"status": "ok"} 200
• curl http://localhost:5000/api/hello → {"message": "Hello, World!"} 200
• Frontend verified loading correctly at http://localhost:3000

Level 4 — volume persistence

• docker volume ls confirms unsdg-classifier-tool_hf-model-cache present
• After docker compose down && docker compose up, backend reached healthy state instantly with no model re-download

---

Notes

• Pairs naturally with Configuration is hardcoded instead of using environment variables #29 (env var refactor) — the .env file Docker needs is the same one that issue introduces
• No classification logic was changed in this PR
• The bash.sh workflow continues to work exactly as before for contributors who prefer not to use Docker