v0.16.0
Highlights since v0.15.3:
- New visible-mark detector: Qwen (千问AI生成), strict-only, calibrated on a 117-frame cohort
- Removed the
assume_aisensitivity mode;--sensitivityis nowauto/strictonly (migration raises loudly) - Fix: faint text marks detected by the tophat front-end are now actually removed (mask from the detector's match box; was a silent no-op on ~8% of Doubao detections)
- Fix: failed output writes now raise a clear error instead of crashing on a downstream stat
- Fix: passing a directory as a single-image source is rejected at parse time instead of IsADirectoryError
- Tophat detection front-end for Doubao, rival-margin gating, Qwen cohort calibration
- Deps: setuptools 81.0.0 -> 83.0.0 (PYSEC-2026-3447); torch 2.13.0 in the lock closes GHSA-rrmf-rvhw-rf47