[bghira]

MeanFlow divided every finite-difference derivative by the fused guidance scale, including dropped-caption samples whose conditional and unconditional predictions share the same model and conditioning. This changes the target of an already-correct unconditional flow.

Carry a per-sample conditioning mask from collation and scale derivatives only where guidance applies. A model foundation hook lets Ideogram retain guidance against its separate unconditional transformer. Tests cover selected and dropped captions, mixed batches, dedicated unconditional models, invalid masks, and exact Gaussian-flow fixed points. The new AnyFlow regressions participate in standard unittest discovery.

Validation:

• .venv/bin/python -m unittest -v -f tests.test_anyflow_dropout_guidance tests.helpers.distillation.test_anyflow_distiller tests.test_collate tests.test_ideogram_unconditional — 92 tests passed.
• The mixed-caption and exact-flow regressions fail against the original guidance calculation.
• All pre-commit hooks passed.
• Full unittest discovery passed all tests before a fixed-port webhook fixture collision between parallel suites. The webhook fixture and remaining 479 tests passed when rerun serially.