[bghira]

Add --adapter_model_precision=no_change|fp32 to the CLI, environment configuration, and WebUI. FP32 preserves small trainable LoRA/LyCORIS and auxiliary parameter updates while frozen base weights retain their configured precision. The default preserves existing behavior; FP32 rejects the BF16-only optimizer.

Upcast before adapter loading and optimizer construction, retain precision through device moves and LyCORIS checkpoint restore, and document the memory tradeoff in every OPTIONS translation. Strengthen MOSS tests to update the dT-MLP before verifying inference, initialization, and checkpoint reloads.

Validation:

• .venv/bin/python -m unittest -v -f tests.test_adapter_model_precision tests.test_moss_image — 47 tests 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.
• All pre-commit hooks passed.