Root cause: 3D RoPE precomputed for max_frame_num+100=1124 positions.
GX019817 has 1357 frames after trim → index overflow → tensor mismatch.
2048 → supports up to 2148 frames (covers all current segments).
Co-Authored-By: Poulpe <claude@nowyouknow.fr>
- kill_stale_demo_py() before each segment to prevent GPU contention from orphan processes
- Remote script runs demo.py in background via nohup, polls for PLY file every 30s, kills viser server once PLY written — prevents indefinite SSH block on viser listener
- offload_to_cpu now read from thresholds.yaml[inference] (default false for 24GB VRAM)
- timeout reads inference_timeout_s from yaml (already 10800s)
- min_frames guard included (from fix/05-inference-min-frames-timeout)
Root cause: demo.py starts viser server after writing PLY; SSH timed out → orphan; two orphans competed for GPU with offload_to_cpu → pure CPU inference = 6h+ for 493 frames
- Ajoute _load_inference_cfg() qui lit config/thresholds.yaml
- Mode/conf/keyframe_interval/max_frame_num depuis config (streaming par défaut)
- Valide par GX049839_v2: streaming+conf=1.5+kf=1 → 146M pts vs 0 pts en windowed sans conf_threshold
- Ajoute --offload_to_cpu (stable sur RTX 3090 .84)
Stage 04 frame extract:
- New lib_frame_qc.py: per-frame Laplacian/contrast/blue-dominance scoring
- Classes: bottom_visible / water_no_bottom / turbid_water / out_of_water
- Sample 1/5 frames after extraction, write qc.json per segment
- Record metrics (frames_total, frames_bottom_visible, bottom_visible_pct)
- Mark job degraded when bottom_visible_pct < 50%
Per-AUV viser view:
- scripts/viser_auv.py loads all PLYs of an AUV, color per file
- POST /pipeline/missions/{id}/auvs/{auv}/view rsyncs ply -> worker
- launches viser on hashed port 9300+, returns URL
- _pipeline.html exposes AUV list, JS handler opens viser tab
- db(): timeout=30 + PRAGMA busy_timeout=5000 pour éviter locked sous charge parallèle
- _ssh_via(): helper routant SSH z620 via worker (.82→worker→z620)
- _extract_on_remote_host(): toutes les commandes z620 passent via worker
- resume probe: ffprobe z620 également via worker
- reset jobs 11/17/18/20 en queued pour relancer
- dispatcher: col step ajoutee (migration). set a chaque phase:
scp N/M, ffmpeg N/M, trimming hors-eau, reconstruct demo.py
- _refresh_thumbnail() scp la DERNIERE frame extraite toutes les ~15s
pendant ffmpeg pour que le preview colle a la progression live
- template: cache-bust thumbnail via ?t=mtime, step affiche sous
progress bar, thumb revenue a 48x27
- CSS: .badge.busy -> animation spin (rotation infinie) au lieu de
juste une couleur, .step-text italique mute
