[TechEngineBot]

What & why

Part of #98. A CPU record with tdp_w: 11.5 passed validation and only failed later, when the dump built the read schema, because the API types the field as an integer. The validator now rejects fractional numbers, booleans and strings in every field the models type as an integer, so the dataset's PR checks catch it before a dump does.

Changes

• app/validate.py: explicit sets of integer fields per category, next to the required-field sets, and one shared check. Whole floats such as 65.0 stay valid: the serialiser accepts them and 15 existing CPU records hold one. The device catalog's own integer check is replaced by the shared one.
• Standard library only: the validator still imports nothing outside the standard library, app.categories and app.data_root. TechAPI's PR checks run it from the submodule without installing the engine, so this matters. A unit test imports it in a subprocess with -I -S.
• Tests: a test derives the integer fields from the models and fails when a model gains an integer column the validator's sets lack; another checks that no read schema types a field as an integer unless the model does.

Checks

• Full TechAPI dataset with the system Python and nothing installed: passes, zero errors, run time unchanged (about 25 s).
• ruff, mypy app, pytest tests/unit (698 passed) locally.