[MateoLostanlen]

• While an alert is ongoing, the engine sends the last 11 frames of the pose, with their boxes, to a new pyro_temporal_api service running the temporal smoke model. The verdict comes back on the next round. Each box is uploaded with its tube score (temporal_model_scores, feat: accept the temporal score computed on the device pyro-api#723), the API applies the threshold. Boxes outside every tube, or in a tube scored 0.2 or less, are dropped.
• Fail open: if the service errors, loses the job or takes more than 120 s, the frames go out without scores and the API scores them itself. An event that ends while a job is pending waits for the verdict, so the last alert of an event is never lost.
• Each distinct frame is sent once. The camera API can serve the same capture twice, and a repeated frame turned a 0.003 false positive into 0.75.
• The service is the torch-free ONNX runtime from temporal-model (feat: ONNX export and torch-free runtime for the tube classifier temporal-model#76, pinned to 7c1500e until a release ships model_onnx.zip), behind the compose profile temporal. Without TEMPORAL_API_URL nothing changes. make run-no-temporal starts the stack without it.
• The service only listens on loopback and caps its inputs: 11 frames of 4 MB per job, 50 boxes per frame, 16 jobs pending. /health returns 503 when the worker is dead or stuck.
• model_onnx.zip must come from the merged temporal-model code: dynamic batch axis, and the format 2 split backbone/head lets the service cache ViT embeddings across jobs.
• Engine and camera API images move to bookworm, bullseye's security repo no longer resolves.
• Temporal API tests run in CI and make test.

Tests

• Pi 5, 8 dataset sequences (5 smokes, 3 non-smoke) replayed through the real engine: every smoke was uploaded after a positive verdict, no non-smoke reached the gate, 0 errors. About 9 s per job, 530 MB RSS for the service, host peaks at 2.4 GB of 4 GB.
• pyro-envdev with fake cameras (feat: fake Reolink camera driven by scenarios pyro-envdev#65) and pyro-api#723: the videlles-01 false positive is rejected (max 0.12), the real fire of smoke-then-clear is still validated (0.97 to 1.0). On the 376 val sequences, engine boxes give the same verdicts as the model's own detector.

[codecov]

Codecov Report

❌ Patch coverage is 93.54839% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 83.02%. Comparing base (4ba8abb) to head (3193c75).
⚠️ Report is 3 commits behind head on develop.

Files with missing lines	Patch %	Lines
pyroengine/temporal_client.py	57.89%	8 Missing ⚠️
pyroengine/engine.py	98.52%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #415      +/-   ##
===========================================
+ Coverage    80.51%   83.02%   +2.50%     
===========================================
  Files            6        7       +1     
  Lines          847      978     +131     
===========================================
+ Hits           682      812     +130     
- Misses         165      166       +1     

Flag	Coverage Δ
unittests	83.02% <93.54%> (+2.50%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.