Scrolls workbench 0.3.0

Workbench 0.3.0 adds explicit reader weight profiles, local multilingual OCR and validated runtime switching.

• Greek reads with supplied line positions now default to letters-v2-real (CC BY-NC 4.0). Automatic region finding keeps letters-v2-ccby (CC BY 4.0). Explicit --weights or --weight-profile overrides remain available.
• --backend tesseract --language lat supports Latin with an installed language pack; other installed recognition languages are supported too. Output records pack hashes, word confidence and blank/shuffle controls.
• ReaderSession.switch() validates a replacement before activating it and preserves the working model on failure.
• Segmentation dataset preparation rejects ambiguous files, malformed metadata and invalid fold counts before writing plans.

Install the tagged version in a virtual environment separate from AutoPipeline:

python -m pip install "scrolls-workbench[reader] @ git+https://github.com/ChaseHendrick/Scrolls@workbench-v0.3.0"

Tesseract and its language packs are optional system dependencies and must be installed explicitly. Existing Greek checkpoint assets and tags are unchanged. Printed OCR examples do not establish accuracy on scrolls; one Latin shuffle control outscored its input. Every proposed transcription remains model output for human review.

Validation: 1,056 tests passed with 61 optional skips, plus 45 numerical checks. Independent review exercised Greek/Latin switching and seven failed-switch cases. The installed wheel passed version, import and OCR smoke checks.