DSpace: optical character recognition for image-only deposits

Deferred in the implementation plan. A deposit whose files are scanned images has no extracted-text bundle, so it is indexed from its description alone and the build says so. OCR would change that, but it means a large dependency and a malformed-input surface the project has avoided everywhere else. If it is ever built, run it on the operator's machine only, never on a scheduled runner, and keep it behind an extra.

Start by reviewing the implementation plan and the current build behavior for image-only deposits, which are indexed from their description alone. Define the OCR approach and its malformed-input handling before implementation. Done means OCR runs only on the operator's machine, never on a scheduled runner, is behind an extra, and makes extracted text available to indexing.

Written by the indexing model from the issue text.