fix(llm): avoid mutating caller messages

[iroiro147]

Summary

• clone caller-provided message lists before public generate_response prompt preparation
• cover base, OpenAI/OpenAI-compatible, Anthropic, Gemini, and GLiNER2 response paths
• add regressions proving reused caller messages do not accumulate schema/language/preamble/retry prompt mutations

Closes #1629

Tests

• uv run pytest tests/llm_client/test_client.py tests/llm_client/test_openai_client.py tests/llm_client/test_openai_generic_client.py tests/llm_client/test_anthropic_client.py tests/llm_client/test_gemini_client.py -q
  • 72 passed, 1 warning
• make lint
  • ruff passed
  • pyright passed with 0 errors
• make test
  • environment-blocked after 171 passed / 2 skipped / 10 deselected
  • failed in Neo4j-backed tests/test_add_triplet.py setup because no local Neo4j service was available at localhost:7687; interrupted after repeated neo4j.exceptions.ServiceUnavailable connection retries

[zep-cla-assistant]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[iroiro147]

I have read the CLA Document and I hereby sign the CLA behalf on myself, e-mail: sarthak.singh@juspay.in

[iroiro147]

recheck

[iroiro147]

Hi maintainers — a gentle follow-up on this bug fix for #1629. It keeps caller-provided messages immutable across the base and provider-specific response paths, with focused regression coverage. I refreshed it onto current main; CLA, Ruff, and fork checks pass, and the remaining triage failure is the repository automation’s external-contributor permission check. Happy to adjust the approach or scope if you would prefer anything different.