test: reorganize suite by module under test and remove duplicates

[vboussot]

Summary

Reorganize tests/unit from per-bugfix files into per-module files, without adding or losing any test behaviour: 60 → 38 unit files, net −460 lines, per-module coverage byte-identical before/after.

What changed

• New tests/conftest.py: an autouse fixture sets the mandatory KONFAI_config_file / KONFAI_CONFIG_MODE defaults via monkeypatch.setenv (per AGENTS.md §7), plus shared write_config and image_attributes fixtures. This removes the 12 copy-pasted module-level os.environ.setdefault blocks and their # noqa: E402 trails.
• Tests regrouped by the unit they cover: the audit/fixes grab-bags (test_audit_fixes.py, test_*_fixes.py, test_*_audit.py, test_data_pipeline_audit.py) and single-test files are redistributed into test_config.py, test_patching.py, test_data_manager.py, test_transform.py, test_augmentation.py, test_network.py, test_blocks.py, test_schedulers.py, test_models.py, test_dataset.py, test_measure.py, test_main_cli.py. Regression comments and AUDIT.md references are preserved.
• 5 exact duplicates merged into their surviving twin, keeping every unique assertion:
  • config missing-file error (path + mode + hint checks folded into one test)
  • invalid-YAML error (×2 → 1, "Invalid YAML syntax" assertion kept)
  • invalid-Literal error (options-mentioned assertion kept)
  • Data._split tail dedup (no-duplication/subset assertions folded into the shard-balance test)

Verification

• Nominal accounting: pytest-collected test IDs before (340) vs after (335) — the diff is exactly the 5 intended duplicates; nothing lost, nothing double-counted, nothing renamed.
• Coverage proof: both trees run against the same konfai source in the same env — the --cov=konfai table is identical line-for-line, module-by-module (8844 stmts, 3254 miss, 63%), proving the removed tests covered no unique line.
• Full suite green: 338 passed (unit + integration) in ~44 s; ruff check/format clean on every touched file (pre-existing violations in untouched files left as-is).

No test was added and none weakened; built-in measures/models intentionally stay untested as agreed.

Follow-up commits

• test: merge the trainer-related files (test_trainer_checkpoints / test_early_stopping / test_ema_convention / test_resume_lr_override / test_resume_https_checkpoint → test_trainer.py) and the runtime files (test_runtime_guards / test_runtime_progress_ddp → test_runtime.py) — all 31 tests moved verbatim, collected IDs unchanged; fix every remaining ruff violation in tests/ and wire pytest.mark.integration so -m selection works.
• ci: gate the core test suite in ruff: the CI lint/format jobs and both pixi task blocks now target konfai konfai-apps/konfai_apps tests.

Final state: 36 unit files (was 60), ruff check/format --check clean over the whole repo including tests, 338 tests passing.