Add PaddleOCR-VL-1.5 to the model registry

[davanstrien]

What

Adds PaddleOCR-VL-1.5 (PaddlePaddle/PaddleOCR-VL-1.5, 0.9B) to MODEL_REGISTRY, completing the PaddleOCR-VL pair alongside 1.6 (added in #21).

Why

Follow-up to #21, which added 1.6. Unlike 1.6 (Qwen3.5/flashinfer → needs the prebuilt vllm/vllm-openai image on a100), the 1.5 script uses transformers batch inference (no vLLM/flashinfer), so it's a standard drop-in: default uv-script image on l4x1. #21 itself flagged 1.5 as an easy drop-in follow-up.

Changes

• run.py — register paddleocr-vl-1.5 as a standard model (l4x1, no image/python/env). Script default --task-mode is ocr (markdown), directly comparable to the other OCR models.
• test_run.py — registry count 9→10; guard that 1.5 is standard (l4x1, no image-mode kwargs).

Verification

• uv run ruff check src/ tests/ — clean
• uv run --extra viewer pytest tests/ -q — 253 passed
• ocr-bench run … --list-models → paddleocr-vl-1.5 shows l4x1 (not image-mode)
• ocr-bench run … --models paddleocr-vl-1.5 paddleocr-vl-1.6 --dry-run → 1.5 standard image, 1.6 image-mode
• Smoke test on HF Jobs against davanstrien/encyclopaedia-britannica-1771 — results added as a comment below

🤖 Generated with Claude Code

[davanstrien]

Closing — smoke-tested on HF Jobs against davanstrien/encyclopaedia-britannica-1771 and PaddleOCR-VL-1.5 errors on every page:

[OCR ERROR: 'PaddleOCRVLImageProcessor' object has no attribute 'min_pixels']

The paddleocr-vl-1.5.py uv-script (uv-scripts/ocr) reads processor.image_processor.min_pixels, which transformers ≥5.0 removed (the script pins transformers>=5.0.0). The registry wiring in this PR is correct, but the underlying model script is non-functional on the current image, so adding it would only record OCR errors.

PaddleOCR-VL-1.6 (merged in #21) is verified working on Jobs and covers PaddleOCR in the bench. Can revisit 1.5 if/when the script is fixed upstream.