[yeongseon]

Summary

Avoid importing the legacy Pydantic namespace solely to reject legacy models. A legacy class cannot exist without its module already being loaded, so consult sys.modules and retain real subclass recognition rather than module-name guessing.

Changes

• Preserve informative v1 rejection for all six model options and subclasses.
• Add a fresh-process regression proving normal v2 decoration leaves pydantic.v1 unloaded; it failed before this fix.

Validation

• Python 3.11 make check-all build: 764 passed, 27 deselected, 4 warnings in 19.51s; 97.16% coverage.
• Existing v1 rejection, custom adapter and v2 behavior tests pass unchanged.
• Repeatable import-cost probe: hatch run python -X importtime -c "import pydantic.v1". On Python 3.11.15/Pydantic 2.14.1 this measured 58.676 ms cumulative including 35.388 ms for pydantic itself (about 23.3 ms incremental legacy import). This is one local observation, not a cross-version performance guarantee; GitHub matrix verifies behavior on supported versions.
• No new dependencies or public options. GitHub CI pending.

Related Issues

Closes #594

Merge authorized by the maintainer’s explicit instruction.

[yeongseon]

Merge authorized by the maintainer’s explicit instruction. Reviewed sys.modules-based identity against existing v1 subclasses and v2 cold-import regression. Exact head b1204db: all five required contexts completed=success, no unresolved threads. Python 3.11 local make check-all build: 764 passed, 27 deselected, 4 warnings in 19.51s; 97.16% coverage.