fix(sdk): normalize Responses tool call IDs

[zora-zi]

HUMAN:

I reviewed the changes and test results, and confirmed that incompatible tool-call IDs are normalized while function calls remain paired with their outputs.

---

AGENT:

Why

Fixes #4578. Tool-call IDs from another provider may be incompatible with the
OpenAI Responses API when conversation history is replayed.

Summary

• Deterministically encode incompatible IDs for Responses.
• Use the same encoded call_id for each function call and its output.
• Preserve compatible IDs and keep parallel calls distinct.

Issue Number

#4578

How to Test

The regression test replays parallel colon-delimited IDs through
LLM.format_messages_for_responses and verifies that calls and outputs remain
paired.

• Targeted tests: 25 passed
• LLM tests: 984 passed
• Pre-commit checks: passed

Video/Screenshots

Not applicable.

Design Doc

Not applicable.

Type

• Bug fix
• Feature
• Refactor
• Breaking change
• Docs / chore

Notes

The encoding uses a provider-specific wire-ID policy. A broader cross-provider
identity design is being drafted.

[enyst]

Hey @zora-zi , thank you for the PR!

In the issue, a comment says they cannot reproduce with the same LLMs, Kimi and GPT-5.6. How about you, did you reproduce it?

Separately, the PR uses an algo to make it very different than the original string, which I’m not sure it’s necessary and might cause other problems. Could we keep it to just the issue detected : -> _, if we can replicate it?