[MateoLostanlen]

• While an alert is ongoing, the engine scores the last 11 frames of the camera pose with the temporal smoke model every round (sends them with their boxes to pyro_temporal_api, reads the verdict on the next round) and uploads each box with the probability of the tube covering it (temporal_model_scores, feat: accept the temporal score computed on the device pyro-api#723, which applies its own threshold instead of scoring the sequence). Boxes below 0.1 or outside every scored tube are dropped; only the frames a verdict scored are uploaded on it, and frames of an ended event the model rejected are dropped for good. A service error, a lost job or a job pending longer than 120 s fails open: the frames go out without scores and the API scores them itself. An event that ends while its job is still pending is kept until the verdict is in, so the last alert of an event is never dropped.
• The service runs the torch-free ONNX runtime from temporal-model (feat: ONNX export and torch-free runtime for the tube classifier temporal-model#76, merged, pinned to commit 7c1500e until a release ships model_onnx.zip), behind the compose profile temporal. Unset TEMPORAL_API_URL keeps today's behaviour. make run-no-temporal starts the stack without it whatever .env says.
• The service binds to loopback and bounds what it accepts: at most 11 frames of 4 MB per job, 50 boxes per frame, 16 jobs pending, 100 results kept. Boxes must be finite, normalized xyxy; frames must be valid images with unique ids. /health returns 503 when the scoring worker died or is stuck.
• model_onnx.zip must be exported from the merged code: the graph now has a dynamic batch axis and older exports fail with more than one tube. A format 2 package (split backbone/head) lets the service cache ViT embeddings across jobs.
• Engine and camera API images move to bookworm since bullseye's security repo no longer resolves.
• Temporal API tests run in CI and make test.

Tested on a Pi 5 by replaying 8 dataset sequences (5 smokes, 3 non-smoke) through the real engine and the service, one per camera pose, with the format 2 0.6.0 package. Every smoke was held one round and uploaded on the next after a positive verdict, the 3 non-smoke sequences never reached the gate, 0 errors. Scoring takes about 9 s per 10-frame job on 4 cores, the service uses about 530 MB of RSS and the host peaks at 2.4 GB of 4 GB with the engine running.

[codecov]

Codecov Report

❌ Patch coverage is 93.10345% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.85%. Comparing base (4ba8abb) to head (3c70979).

Files with missing lines	Patch %	Lines
pyroengine/temporal_client.py	57.89%	8 Missing ⚠️
pyroengine/engine.py	98.41%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #415      +/-   ##
===========================================
+ Coverage    80.51%   82.85%   +2.33%     
===========================================
  Files            6        7       +1     
  Lines          847      968     +121     
===========================================
+ Hits           682      802     +120     
- Misses         165      166       +1     

Flag	Coverage Δ
unittests	82.85% <93.10%> (+2.33%)	⬆️

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.