Phase 2 groundwork: Jetson detection, honest power, ESP32-S3 CNN inference

[tarun101]

Follows up the Jetson/Wyze/ESP32 camera support with the pieces that make those platforms actually usable for Phase 2 (Orin Nano + ESP-CAM). Three independent commits.

1. Detection on other cameras — calibration profiles (46ba05f)

Box-finding + warp were hard-coded to one imx708 capture, so detection couldn't locate the grid on any other camera. Lifted into a camera profile (detect/profiles/<name>.json), selected by PILLBOX_CAMERA_PROFILE (default imx708).

• camera_profiles.py loads/validates the active profile; crop_cells.py exposes its fields under the same module names, so every caller is untouched.
• profiles/imx708.json reproduces the original constants bit-for-bit (asserted) — the Pi deployment and all committed pillbox-data numbers are unchanged.
• calibrate_camera.py generates a profile for a new camera from one empty-box photo: mark 4 corners (CLI --quad or interactive --click), anchors derived automatically.
• To bring up a Jetson/Wyze cam: photograph the empty box, run the tool, set PILLBOX_CAMERA_PROFILE, and DoG/CNN/YOLO all work against it.

2. Power measurement — honest Orin Nano path (9e19b12)

The Jetson tegrastats reader returns None on the Orin Nano/NX (those modules drop the onboard INA3221 sensors), so the paper would silently get no power there.

• Adds an external-meter source (PILLBOX_POWER_METER_CMD: any command that prints watts — USB meter, smart plug, bench PSU) as the highest-trust reader — the real way to get board power on the Orin Nano.
• Pi 5 PMIC and Jetson tegrastats paths unchanged; paper_stats now reports the method used and tells the operator to set the meter var on a Jetson with no rails instead of implying power was captured.

3. ESP32-S3 on-device CNN inference (8e91125)

Phase-2 goal of running detectors on the board itself.

• convert_cnn_to_tflite.py: pill_classifier.onnx → int8 TFLite → C header, int8-calibrated on real 6-channel cells.
• main/main.cpp: esp-tflite-micro harness — runs the CNN on-device, reports latency / arena usage / prediction over serial.
• ESP-IDF scaffolding (pinned esp-tflite-micro, PSRAM config) + README with a model-fit analysis measured from the shipped ONNX: Ref-CNN (79.6k params, ~78 KB int8) is the primary fit; DoG ports to C; YOLO (1.44 M) fits PSRAM but is a slow stretch goal.

Validation status

• Tested here (40 unit tests, all pass): profile load/validate, anchor derivation, calibrate tool, PMIC/tegrastats/external-meter parsing + source priority, the conversion script's representative-set + C-header logic. imx708 profile asserted to reproduce originals; DoG pipeline smoke-tested end-to-end.
• Needs hardware to validate (called out in code/README): the ESP32-S3 firmware (compile + arena/op tuning) and the TF-side ONNX→TFLite conversion. Written to standard APIs, not run on a board here.

🤖 Generated with Claude Code

---

Generated by Claude Code