v0.12.0

Changes since v0.11.4

New pipelines and flags:

• feat(invisible): --pipeline qwen -- Qwen-Image 20B MMDiT img2img pipeline (Apache-2.0, CUDA/cloud-class), best text preservation including CJK. Manual opt-in, no auto-router.
• feat(invisible): --tile -- sliding-window tiled diffusion for large inputs, the lossless alternative to a --max-resolution downscale on MPS/GPU OOM. Feather-blended tiles near SDXL's native size; engages only when the long side exceeds --tile-size.
• feat(invisible): region-targeted regeneration for AI-enhanced composites -- regenerate only the AI box and feather-composite it back, leaving the real photo pixel-exact.

Detection and metadata:

• feat(identify): broader C2PA vendor coverage, a generated/enhanced digital-source-type split, and unified detect/remove thresholds.
• feat(metadata): blank AI-generator tokens in AVIF/HEIF Exif meta-box items in place (piexif-validated, camera/editor EXIF left intact).

Fixes:

• fix(visible): over-subtraction guard for the Doubao/Jimeng/Samsung text marks -- removes the dark-pit ghost left by reverse-alpha on dark and mid-tone backgrounds.
• fix(qwen): native-geometry img2img (no more 1024x1024 squish of non-square inputs) and pipeline-aware strength so --pipeline qwen carries its own certified floors.