Runaway whitespace fix

[will-exaforce]

This PR improves the output generated by the model and the model output normalization when llm is enabled.

This was part of an effort to address issues I observed while validating SkillSpector LLM support locally.

All tests pass locally. I'll get another PR up later to add the proper workflow. In the meantime:

627 passed, 12 skipped, 26 deselected, 1 warning in 1.86s

The benchmark tool was used to compare accuracy and performance against main

Finding summary

Comparison performed across 50 (or all if < 50) samples of each type of data set in https://github.com/lxyeternal/MalSkillBench/tree/main/Dataset.

Performance

runaway-fix completed ~30% faster
runaway-fix - 35:07
main - 46:54

Errors

runaway-fix:

• None

main:

• 15 (All timeout)
• Impacted files
  • 7 of the 50 sampled malicious skills
  • 4 of the 50 sampled malicious python json files

False Negatives

Unchanged across PRs

False Positives

runaway-fix (5):

• Skills/benign/4chan-reader
• Skills/benign/abn-skill
• Skills/benign/aaveclaw
• Skills/benign/aap-passport
• Skills/benign/2captcha

main (1):

• Skills/benign/abn-skill