Summary

Removes hidden runtime environment mutation from optional SDK paths.

• Replaces Spark/PySpark runtime pip install helpers with clear optional-extra errors.
• Stops spaCy annotators from auto-downloading models during normal SDK use; missing models now point to explicit datafog download-model ... guidance.
• Loads GLiNER models with local_files_only=True during normal SDK use, while preserving the explicit CLI download path.
• Stops Donut OCR from installing torch / transformers at runtime and loads Donut model artifacts from local cache only.
• Adds pyspark to the distributed extra so Spark install guidance is accurate.
• Updates CI/release setup to explicitly download the GLiNER model before nlp-advanced tests.
• Adds regression tests covering no runtime package installs and missing optional dependency errors.

Validation

• .venv313/bin/python -m black --check ... passed for changed Python files.
• .venv313/bin/python -m pytest tests/test_runtime_dependency_safety.py tests/test_donut_lazy_import.py tests/test_gliner_annotator.py tests/test_engine_api.py tests/test_client.py -q passed: 55 passed, 2 expected fallback warnings.
• .venv313/bin/python -m pytest tests/ -m "not slow" --ignore=tests/test_image_service.py --ignore=tests/test_ocr_integration.py --ignore=tests/test_spark_integration.py --ignore=tests/test_text_service_integration.py -q passed: 516 passed, 10 skipped, 291 deselected, 19 xfailed.
• .venv313/bin/python setup.py egg_info passed, with existing setuptools license-classifier warning.
• .venv313/bin/python -m pip install --dry-run -e ".[distributed]" resolved and showed pyspark included in the distributed extra.
• .venv313/bin/python -m pytest tests/test_spark_integration.py -q skipped cleanly without installing PySpark.
• .venv313/bin/python -m pip check passed.
• git diff --check passed.

Note: tests/test_ocr_integration.py cannot collect in this core-only venv because PIL is intentionally absent; OCR integration remains an optional-profile test.