Document and test-isolate GRAPHTAGE_BATCH_BACKEND

[ESultanik]

GRAPHTAGE_BATCH_BACKEND pins the batch Levenshtein backend that #195 introduced. Two things about it need
attention before it is named in the 0.5.0 release notes.

The variable broke the test suite for anyone who used it

Two tests assert that a batch reaches the numpy backend when nothing pins a backend. Both read the ambient
environment, so exporting the variable made them fail:

$ GRAPHTAGE_BATCH_BACKEND=python pytest -q
FAILED test/test_batch_distance.py::TestBackendSelection::test_environment_variable_overrides_the_automatic_choice
FAILED test/test_batch_distance.py::TestCostOracle::test_one_big_pair_goes_to_a_vectorized_backend
2 failed, 270 passed

The product was correct in both cases — the variable did exactly what it promises. The tests simply could not
observe the unpinned baseline they were written to check, so anyone setting the variable to exercise the other
backend saw two spurious failures from tests that never mention it.

A new no_backend_pin() helper removes the variable for the span that watches the automatic choice, leaving the
halves that pin a backend explicitly alone.

It was documented only in a module docstring

rg GRAPHTAGE_BATCH_BACKEND hit only graphtage/batch_distance.py and its test. The README now carries an
"Environment Variables" section alongside the rest of the command line surface.

Testing

Both tests were checked against deliberate breaks rather than assumed to still bite:

Mutation	Result
Ignore the variable at both read sites in batch_distance.py	test_environment_variable_overrides_the_automatic_choice fails
Raise VECTORIZED_MIN_CELLS to 10^9	test_one_big_pair_goes_to_a_vectorized_backend fails

The suite passes with the variable set to python, set to numpy, and unset — 272 passed, 1770 subtests, in each
case. Ruff clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01F2sHz5c5TvMs9tFn2HhwaC