Motivation

• Hosted notebook platforms (Google Colab, Kaggle) expose secrets via platform secret stores rather than the process environment, preventing Hugging Face libraries from seeing HF_TOKEN during import.
• Populate the conventional HF_TOKEN from notebook secret stores at startup so gated model access works in hosted notebooks without manual environment setup.
• Preserve any explicitly configured HF_TOKEN and treat missing or inaccessible notebook secrets as non-fatal to avoid breaking anonymous access.

Description

• Add unsloth/notebook_token.py which detects the default secret name HF_TOKEN, reads it from Colab (google.colab.userdata) and Kaggle (kaggle_secrets.UserSecretsClient), strips whitespace, and sets os.environ['HF_TOKEN'] when found.
• Invoke the detection early from unsloth/__init__.py so the token is available before Hugging Face libraries freeze their authentication configuration.
• Add tests/test_notebook_token.py with tests for Colab detection, Kaggle fallback after a missing Colab secret, explicit HF_TOKEN precedence, and non-fatal handling of secret-access failures.

Testing

• Ran pytest -q tests/test_notebook_token.py and all tests passed (4 passed).
• Compiled the new module with python -m compileall unsloth/notebook_token.py and it succeeded.

Codex Task