test_metadata_reader.py previously only tested the private
_lookup_metadata helper, never the node's actual read() method -- so the
README's "full unit-test coverage" claim for this node was true of the
helper but not the public interface. Adds two tests that call read()
directly: one demonstrating the real-world limitation the README already
documents (a real IMAGE tensor always loses its metadata through the
tensor-to-PIL conversion, regardless of embed_key), and one confirming
read() degrades gracefully on bad input instead of raising.
Building a real tensor requires torch, so this moves
test_metadata_reader.py from the torch-free to the torch-required group
in CONTRIBUTING.md and the CI workflow (5/9 split now, was 6/8). The new
split is empirically verified by running all 14 tests with torch's
import blocked, not just inferred from grep.
pyproject.toml's pending dependencies fix is intentionally NOT included
in this commit -- it would re-trigger the publish workflow against the
unchanged "1.1.0" version string. That's bundled into the version-bump
commit instead.
SurrealByDesign
e266dc5218