TASK 30 — Raising the detection floor to 0.6 and gating the seam merge

Why

Reviewing the carpet fragments frame by frame ([[TASK 29]]) surfaced a real bug in where the score threshold was applied. The panorama face-seam merge (merge_seam_duplicates, TASK 25) runs inside the sidecar, right after YOLO detects each face. But the offline benchmark froze detections at a 0.25 floor and re-filtered to a higher floor after the fact — which is after the seam merge already ran. So a 0.9-confidence carpet detection could be unioned with a 0.3-confidence fragment, inheriting the high score and carrying the fragment's mask contamination past every offline filter. arabic_room vp_002 showed exactly such a spurious carpet.

What changed

• Score threshold raised to 0.6 across the stack. The per-face gate that matters was always there — it is YOLO's conf=score_threshold in PerceptionPipeline.detect (perception/pipeline.py), applied before SAM and before the seam merge. Raising the floor to 0.6 means low-confidence fragments never enter the seam merge. Defaults updated in: responder.py (DEFAULT_SCORE_THRESHOLD, the live value), pipeline.detect, client.detect, server.py (the /detect Form default), replay.py, perception/__main__.py, and the benchmark tools dump_detections.py, dump_debug.py, replay_score.py, debug_viewpoint.py.
• Frozen detections re-dumped at 0.6. captures_nav/arabic_room was re-run through the sidecar (A10G) at --score-threshold 0.6, so YOLO + SAM + seam merge all operate on the ≥0.6 set. The old 0.25 frozen dets are gone; offline floor sweeps below 0.6 now need a fresh sidecar re-dump.

Result — arabic_room carpet nodes

configuration	carpet nodes	total nodes
0.25 floor	27	199
0.25 dump, filtered to 0.6 offline (after the merge)	13	50
0.6 dump, seam merge run at 0.6	9	46

The step from 13 → 9 is the fix: four carpet fragments existed only because low-confidence detections had been eligible for the seam merge. vp_002's spurious carpet is gone; its one remaining carpet detection lifts inside the GT box. The residual (node 0 + node 26 over GT #58, node 53 + node 14 over GT #32) is cross-frame [[B6 large-object fragmentation]], which the seam merge — a within-frame operation — does not address.

Caveat

Raising the live floor to 0.6 trades recall for precision on small/faint objects; the 0.35 value it replaces was swept for recall. This wants an end-to-end re-score to confirm it is a net win on the challenge metric, not just on node counts. Regenerated debug dump, both merge videos, and the vp_002 / node-0 inspect images all reflect the 0.6 pipeline.