[ishovkun]

Summary

Follow-up to #310. TensorRT-LLM #19504 pins cosmos_guardrail 0.3.2, which copies the guardrail's NLTK data out of the Hugging Face cache as regular files on first use. The server-side NLTK_DATA workaround documented for 0.3.0 is no longer needed.

• Bump the documented guardrail pin from 0.3.0 to 0.3.2 in the shared TensorRT-LLM setup.
• Remove the "Server-side NLTK data setup" section and every link to it from the Action and Transfer READMEs and the three TensorRT-LLM notebooks.
• Update the validated TensorRT-LLM source revision to cc63e66d (1.4.0rc0), which contains #19504 and #17476.

Overlaps with #343 by a few lines in cookbooks/cosmos3/README.md (the guardrail install block); whichever lands second needs a trivial rebase.

[ConstBob]

LGTM. The NLTK workaround section is gone, and the Action/Transfer READMEs and TensorRT-LLM notebooks no longer link to it. The pin is cosmos_guardrail==0.3.2, and the validated TensorRT-LLM revision is cc63e66d. That matches the 0.3.0 symlink failure we hit with guardrails on. I did not re-run 0.3.2 here. Approving from the diff and your B200 check (guardrails on, NLTK_DATA unset).