fix: handle GPT-5 Chat Completions with deterministic parameters

[666ghj]

Summary

• route GPT-5-family Chat Completions requests through a deterministic parameter policy
• use max_completion_tokens and omit unsupported temperature values for GPT-5-family models
• preserve the existing request shape for legacy models
• keep response_format unless the caller explicitly omits it
• extract text safely from supported SDK response content shapes
• propagate provider errors instead of retrying based on error-message text

Why

PR #498 correctly identified that unconditional legacy parameters can make GPT-5-family requests fail. Its adaptive retry removed parameters by matching provider error strings, which could mutate valid requests unpredictably and hide unrelated failures. It also removed supported parameters too broadly and carried unrelated lockfile churn.

The maintained version uses an explicit model-family mapping and keeps provider errors observable.

Validation

• uv run pytest -q tests/test_openai_chat_compat.py — 9 passed
• covered GPT-5 and legacy request shapes, error propagation, and response content extraction
• Python compileall passed
• frontend production build passed
• git diff --check passed
• reconstructed on the current main branch

No paid OpenAI API request was made; the official parameter contract and mocked SDK boundary were used.

Preserves the original contributor commit and supersedes #498.

🤖 Sent by the MiroFish repository maintenance agent.