fix: accept Cerebras tool calls without content

[stackbilt-admin]

Summary

• Allow Cerebras OpenAI-compatible tool-call responses to omit choices[].message.content.
• Keep formatting behavior unchanged: omitted content normalizes to an empty assistant content string.
• Add a regression test for zai-glm-4.7 tool calls where message.content is absent.

Why

AEGIS gateway model-quality eval reached Cerebras after the llm-gateway developer-role compatibility fix, then failed on schema drift:

Response schema drift at choices[0].message.content: expected string-or-null, got undefined

That shape is recoverable and matches the tolerance already used for Groq/Cloudflare OpenAI-compatible tool-call responses.

Validation

• npm run typecheck
• npm test (442/442)

Notes

• Focused provider-contract change only.
• Did not include incidental package-lock version refresh from dependency install.