docs: update CLAUDE.md and README.md for CUDA-only support

Reflect removal of MPS, float16/float32 device logic, corrected
input format, and CUDA-only requirement for slow tests.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

CLAUDE.md

@@ -4,7 +4,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## Project Overview
 
-A Hugging Face Spaces app that translates English text to 183 evaluated languages (Table 14, Section A.9) from Google's [MADLAD-400](https://arxiv.org/pdf/2309.04662) 3B Seq2Seq model. Built with Gradio and deployed on HF Spaces. Falls back to CPU with a warning when no GPU is available.
+A Hugging Face Spaces app that translates English text to 183 evaluated languages (Table 14, Section A.9) from Google's [MADLAD-400](https://arxiv.org/pdf/2309.04662) 3B Seq2Seq model. Built with Gradio and deployed on HF Spaces. Falls back to CPU with a warning when no CUDA GPU is available.
 
 ## Commands
 
@@ -25,21 +25,21 @@ ruff format .
 ty check
 
 # Test
-pytest                     # all tests (slow tests require GPU + model download)
+pytest                     # all tests (slow tests require CUDA + model download)
 pytest -m "not slow"       # fast tests only
-pytest -m slow             # model tests only
+pytest -m slow             # model tests only (CUDA only)
 ```
 
 ## Architecture
 
-**`app.py`** — Single-file application. Uses `@lru_cache` for lazy loading of the `google/madlad400-3b-mt` tokenizer and model with `float16` precision (no download on import). Translation prepends a language token to the input text (e.g., `<2fr>Hello`) before tokenization and generation. The `@spaces.GPU` decorator allocates GPU on HF Spaces infrastructure.
+**`app.py`** — Single-file application. Uses `@lru_cache` for lazy loading of the `google/madlad400-3b-mt` tokenizer and model (no download on import). Uses `float16` on CUDA, `float32` on CPU. MPS is not supported (produces garbage output with T5 models). Translation prepends a language token with a space to the input text (e.g., `<2fr> Hello`) before tokenization and generation. The `@spaces.GPU` decorator allocates GPU on HF Spaces infrastructure.
 
 **`langmap/`** — Package with `langid_mapping.py`, a hand-maintained dictionary mapping 183 evaluated language tokens to human-readable language names (sourced from Table 14, Section A.9 of the MADLAD-400 paper; evaluation sets: WMT, Flores-200, NTREX, Gatones). Available languages at runtime are the intersection of this mapping and the model's vocabulary.
 
-**`tests/`** — Pytest suite split into fast (`test_langmap.py`) and slow (`test_app.py`). Slow tests require GPU and model download; they are auto-skipped without MPS/CUDA. Fast tests in `test_app.py` verify the module imports without triggering model download.
+**`tests/`** — Pytest suite split into fast (`test_langmap.py`) and slow (`test_app.py`). Slow tests require CUDA and model download; they are auto-skipped without CUDA. Fast tests in `test_app.py` verify the module imports without triggering model download.
 
 ## Tooling
 
 - **Ruff** — linter and formatter (`ruff.toml`). Rules: `E`, `F`, `I`, `W`. Line length: 120.
 - **ty** — type checker (`ty.toml`). Python 3.12 target.
-- **pytest** — test runner (`pytest.ini`). Custom `slow` marker for GPU-dependent tests.
+- **pytest** — test runner (`pytest.ini`). Custom `slow` marker for CUDA-dependent tests.

README.md

@@ -18,7 +18,7 @@ Translate English text to 183 evaluated languages (Table 14, Section A.9) using
 ## Requirements
 
 - Python 3.12
-- MPS (Apple Silicon) or CUDA GPU
+- CUDA GPU (recommended) or CPU
 
 ## Setup
 
@@ -38,5 +38,5 @@ ruff check .             # lint
 ruff format .            # format
 ty check                 # type check
 pytest -m "not slow"     # fast tests
-pytest                   # all tests (requires GPU + model download)
+pytest                   # all tests (requires CUDA + model download)
 ```