[joepetjr]

Why

GoMoveShift's blur step calls detect_and_track on a 20 GB, 54 197-frame day video (measured 2026-10-08). The log printed "person model loaded" at 22:40 and then nothing until the mask was written about 4.5 h later (earlier runs: 3h29m and 3h41m of silence). An operator or a stale-run check could not tell "working" from "hung". Checkpoint writes are silent too.

What

• New keyword progress_every_s: float = 60.0 (wall clock, time.monotonic; 0 disables).
• Inside the frame loop: detect: frame N/TOTAL (P%), D detections so far, F fps about once per interval. TOTAL and P are dropped when CAP_PROP_FRAME_COUNT is 0. fps is measured over the interval.
• Final detect: done, N frames in Xs line (frames decoded this call).
• Logging only. Returned boxes unchanged (tested against progress_every_s=0).
• README and CHANGELOG updated.

Evidence

• Red: 4 new tests in TestDetectionProgressHeartbeat failed before the change (detect had no time; the keyword did not exist).
• Green after: 4 passed. Full suite: 271 passed, 0 failed. ruff check .: all checks passed.
• Diff: 4 files, 119 insertions, 0 deletions.