pydantic and numpy were undeclared runtime dependencies (fixed in v2.19.1)

Summary

pydantic and numpy were imported unconditionally at module level in result.py, _validate.py, cypher/translator.py, fusion.py, and vector_utils.py but were not declared in [project.dependencies] — only in the [full] optional extra.

A clean pip install iris-vector-graph==2.19.0 would fail with ModuleNotFoundError unless the user also installed [full].

How it was caught

Architecture review surfaced that 5 packages load unconditionally on from iris_vector_graph import IRISGraphEngine but were absent from declared runtime deps. Verified by tracing the import chain: result.py → cypher/translator.py → pydantic; fusion.py → numpy.

Fix

Added pydantic>=2.0.0 and numpy>=1.24.0 to [project.dependencies] in v2.19.1.

Prevention

Before each release, verify that every unconditional top-level import in the IRISGraphEngine import chain is declared in [project.dependencies]. A quick check:

python3 -c "
import sys
before = set(sys.modules.keys())
from iris_vector_graph import IRISGraphEngine
after = set(sys.modules.keys())
third_party = {m.split('.')[0] for m in after - before
               if not m.startswith('iris_vector_graph')
               and not m.startswith('_')
               and m not in sys.stdlib_module_names}
print('Third-party hard deps:', sorted(third_party))
"

Compare the output against [project.dependencies] in pyproject.toml before every release.

Start with pyproject.toml and inspect [project.dependencies], then run the import-chain check shown in the issue. Compare its third-party modules with the declared dependencies; the reported fix is already present in v2.19.1, so done means confirming the declarations and checking for any remaining unconditional imports.