Faut-il créer votre propre Bevel ?

Build a local Apple Health analytics dashboard. Single machine, no accounts, no cloud, no telemetry.

Stack, non negotiable: Python 3.12, DuckDB for storage, Streamlit for the UI, uv for dependency management. No web framework, no Docker, no database server.

Project layout:
- ingest.py: CLI that takes a path to an Apple Health export.zip
- app.py: Streamlit dashboard
- health.duckdb: local database file, gitignored
- .env for anything configurable (default export path), read with python-dotenv

Ingest requirements:
- Read export.xml directly from inside the zip, streaming, using xml.etree.ElementTree.iterparse and clearing elements as you go. The file can be 500MB+, never load it into memory.
- Extract Record elements and Workout elements. Normalise into two DuckDB tables: records(type, source, unit, start_ts, end_ts, value) and workouts(activity_type, start_ts, end_ts, duration_min, energy_kcal).
- Strip the HK prefixes off type names so they read as sleep_analysis, heart_rate_variability_sdnn, resting_heart_rate, step_count, active_energy_burned, respiratory_rate, oxygen_saturation, body_mass.
- Idempotent: re-running on the same export replaces the tables, does not duplicate.
- Print a summary of row counts per type and the date range found.

Dashboard requirements:
- Date range picker, defaults to last 180 days.
- Daily aggregate table built in SQL: steps sum, active energy sum, resting HR mean, HRV mean, sleep hours from sleep_analysis asleep intervals, workout minutes.
- One chart per metric with a 7 day rolling mean drawn over the raw daily points.
- A correlation panel: pick metric A and metric B, pick a lag of 0 to 3 days, show the Pearson r, the n, and a scatter plot. Print a blunt caption saying this is correlation on a tiny sample and means nothing on its own.
- A weekday breakdown: mean of each metric by day of week.
- No composite "scores", no AI summaries, no advice text. Numbers only.

Deliver a README with the exact export steps from the iPhone Health app, the uv commands to run ingest and the dashboard, and one paragraph stating clearly that this needs a manual re-export to stay current and is not medical advice.