Add model-in-the-loop workflow: vp eval, autolabel, queue, YOLO-seg, semantic masks

[CaioWing]

What

Implements the model-in-the-loop slice of the roadmap (Phase B + Benchmarking + parts of "Later"), plus a few consistency fixes.

New commands

• vp eval — score model predictions against a split set (test by default), turning a locked split + snapshot into a reproducible benchmark. Box tasks get per-class AP@50, mAP@50, mAP@50-95 (COCO-style 101-point interpolation) and precision/recall at a confidence threshold; classification gets accuracy, per-class P/R/F1 and a confusion matrix. --json for machine-readable output; warns when the split is unlocked.
• vp autolabel — persist confident predictions as annotations, recorded with source.type = "model" so model labels stay distinguishable and auditable. Only unlabeled assets are touched unless --replace; --min-confidence filters objects.
• vp queue — active-learning queue: unlabeled images first (ranked by model uncertainty when predictions are given); --include-labeled audits existing labels for ground-truth/prediction disagreement (possible missing or stale labels).

Shared infrastructure

• visionpack/predictions.py — one loader for all three commands. Accepts vp-native JSON, COCO results/instances JSON, and YOLO txt directories (exactly what Ultralytics predict writes with save_txt/save_conf over a vp export layout). Images resolve by asset id or original filename; unmatched references and unknown classes are surfaced, never silently dropped.

Format coverage

• YOLO-seg — imports polygon label lines (class x1 y1 x2 y2 ...) as instance-segmentation geometry; vp export --format yolo writes YOLO-seg labels for segmentation projects (or with --seg); plain boxes degrade to four-corner polygons so mixed datasets stay trainable.
• vp export --format masks — semantic-segmentation export: 8-bit class-index PNG per image (0 = background), split-aware, with a classes.txt documenting the pixel-value mapping.

Fixes

• vp --version now reads the installed package metadata (was hardcoded 0.1.0 while pyproject says 0.0.1).
• Deduplicated the triplicated "Release process" section in the README.
• Refreshed ARCHITECTURE.md (module map, roadmap checkboxes, stale test count), docs/usage.md, and CHANGELOG.

Known limitations (documented in docs/usage.md)

• Segmentation eval uses each polygon's enclosing box; mask IoU is a follow-up.
• YOLO-pose import/export still pending.

Testing

• 18 new tests (test_eval.py, test_model_loop.py, test_segmentation_formats.py); full suite: 118 passing, ruff clean.
• Verified end-to-end with the real CLI: import → split lock → vp eval (mAP 1.0 on echoed GT; precision drops correctly with an injected false positive) → vp queue → vp autolabel → YOLO-seg and masks exports.

🤖 Generated with Claude Code