[Majyles]

Why

As a newcomer, the sample alerts of the dev environment were confusing: footage from distant towers (Aveyron, Chile…) sent to Seine-et-Marne cameras, camera ids in the notebook not matching the seed CSV, unrelated sequences merged into the same alerts. Closes #26 (and goes in the direction of #11).

What

• New scenarios.py module: each scenario replays real footage with the camera that recorded it, so images, camera names and cones on the map agree. Cameras are referenced by name, not id.
• send_real_alerts.ipynb rewritten: a Quick start cell sends every scenario (~12 s), a Start over cell deletes every alert/sequence/detection through the API, then one short cell per scenario explains what to look at in the platform, plus a section to send your own sequence folder with send_sequences().
• Scenarios: a fire triangulated by three towers (2025-07-14, sdis-77), a fire seen by a single camera (Videlles, sdis-77), a fire for sdis-07 (Serre-de-Barre).
• Detections keep the real capture times of the samples, shifted to now, so the alert timeline replays the real pace of the fire without waiting for it.
• Each detection comes with a close-up per box, and each camera gets its last image, so the "close-up" button and the dashboard show real views.
• Same features from a shell: python scenarios.py [name...], python scenarios.py --delete-all-alerts.
• README: documents the scenarios; also fixes the users CSV path and an unclosed code block.

Notes on the sample data

• videlles-02-300: the landscape matches the map at ~120°, not 300° (forest shapes, road landmarks), so the scenario uses 120.
• Brison-03 / Serre-de-Barre (2025-09-29): Brison's box has a 0 confidence, and the triangulated position falls 600 m from Serre-de-Barre while the image shows the smoke several km away (nearly aligned cones). Only Serre-de-Barre is used.
• In selection-true-positives, the prediction files have a 0 confidence, so the ground-truth labels are used instead.
• The samples start with a burst of frames ~1 s apart, then ~30 s: likely buffered frames uploaded at once when the smoke was detected. This leading burst is spread at the typical gap; regular sequences are left untouched.

Tested

Clean database: quick start, start over, full notebook executed end to end, and scenarios.py from a shell. Coherent alerts each time, timelines with one step per frame, close-ups displayed. Pre-commit hooks pass.
Description generated with Claude Code.