feat(tpms): slice-1 ingest + JSON API for Mazda CX-9

[metaframework-dispatch-bot]

Summary

Slice-1 of TPMS ingest for the 2013 Mazda CX-9, per Tom's scoping answers on #7:

• Capture: rtl_433 -M utc -f 315000000 -R 156 -F json → stdout (Abarth-124Spider decoder covers Mazda VDO sensors at 315 MHz per issue Research Task: Tire Pressure Monitoring #1 research).
• Ingest (src/tpms/ingest.py): reads that JSON stream from stdin, filters to the Abarth-124Spider decoder, normalizes pressure to kPa + temperature to °C + battery_ok, upserts into SQLite. Survives malformed lines and non-target decoder events without dropping the pipe.
• API (src/tpms/api.py): Flask app on :8090:
  • GET /api/vehicles — vehicle set (mazda-cx9 only in slice-1).
  • GET /api/vehicles/mazda-cx9/tpms/latest — one row per known sensor, plus any_low / all_stale flags for the future dashboard.
  • GET /api/vehicles/mazda-cx9/tpms/history?since=&until=&limit= — time-window slice.
  • GET /api/health — last_event_ts, receiver_ok, so the dashboard can distinguish sensors-quiet from receiver-broken.
• Data model (src/tpms/db.py): vehicles + sensors + readings. Slice-1 seeds one vehicle; positions stay NULL (Tom's Q4 answer — any-tire-low is sufficient).
• Systemd (deploy/plexpi/tpms-{capture,api}.service) — systemctl --user units targeting the pi account on plexpi. plexpi has no itguy, so deploy/plexpi/README.md has the manual install ritual (bluelinky/fuelbot pattern).

Scope explicitly deferred: sensor→corner mapping, drive-time reception, other vehicles, dashboard UI.

Design notes

• rtl_433 must be invoked with -M utc — without it, timestamps are system-local with no tz suffix and the ingest daemon would skew every reading by the local UTC offset. Captured in both the systemd unit and the module docstring.
• API opens plain sqlite3 connections per request after create_app() runs schema init once. Concurrent reader alongside the ingest writer — WAL is on.
• Duplicate readings drop silently on (ts, sensor_id) PK conflict — rtl_433 sometimes re-emits an identical event.

Test plan

• 43 unit tests pass (21 new + 22 pre-existing). DB idempotency, ingest survives malformed input, non-target vehicle filtering, PSI→kPa fallback, stale + any_low semantics, unknown-vehicle 404, history bounded by since/until, health with empty DB.
• End-to-end smoke: piped a real JSON event into python -m src.tpms.ingest, then hit every API endpoint. See the captured session below.

Captured test run (showboat)

TPMS slice-1 test run

2026-07-13T00:23:41Z by Showboat 0.6.1

uv run pytest

warning: `VIRTUAL_ENV=/Users/tom/Documents/work/metaframework/.venv` does not match the project environment path `.venv` and will be ignored; use `--active` to target the active environment instead
============================= test session starts ==============================
platform darwin -- Python 3.13.7, pytest-9.0.2, pluggy-1.6.0
rootdir: /private/tmp/rf-wt-1783901697-50961
configfile: pyproject.toml
testpaths: tests
plugins: anyio-4.13.0
collected 43 items

tests/test_cli.py .....                                                  [ 11%]
tests/test_device.py .................                                   [ 51%]
tests/tpms/test_api.py .......                                           [ 67%]
tests/tpms/test_db.py ......                                             [ 81%]
tests/tpms/test_ingest.py ........                                       [100%]

============================== 43 passed in 0.18s ==============================

Closes #7

🤖 Generated with Claude Code