[codex] align provider API endpoint contracts

[cafferychen777]

What changed

• Align Qwen and MiniMax with current OpenAI-compatible chat completion endpoints.
• Centralize R OpenAI-compatible provider request, response, and HTTP error handling.
• Normalize R/Python base URL, model, and API key input contracts.
• Add regression coverage for endpoint defaults, URL validation, trimming, and error parsing.

Validation

• python3 -m pytest
• python3 -m ruff check
• Rscript -e 'setwd("R"); devtools::test()'\n- git diff --check\n- ./scripts/check-sensitive.sh\n

[cafferychen777]

Full Code Review

Thoroughly reviewed all 28 changed files, ran all tests, and verified endpoint changes against current provider documentation.

Verified

• Qwen endpoint migration: dashscope-us.aliyuncs.com/compatible-mode/v1/chat/completions confirmed as the current OpenAI-compatible US endpoint via official Alibaba Cloud documentation. DashScope region-specific keys + 3-endpoint fallback logic is correct.
• MiniMax endpoint migration: api.minimax.io/v1/chat/completions confirmed as the current endpoint via official MiniMax platform docs. Legacy .chat and .maxi domains superseded.
• R refactoring: BaseAPIProcessor shared methods (stop_for_http_error, post_chat_completions_request, extract_chat_completions_content, extract_error_message) faithfully reproduce original behavior while eliminating ~500 lines of duplication across 8 processors.
• Input normalization: trimws() / case-insensitive matching at multiple entry points is defensive and correct.
• URL validation: New validate_base_url() properly rejects unsafe schemes, missing hostnames, userinfo, query strings, and fragments.
• Endpoint probing: 401/403/404 now correctly trigger trying next endpoint (region-specific keys return these on wrong host).
• Heuristic JSON repair removal: Correct decision — regex repairs could corrupt valid JSON.
• Backward compatibility: Legacy Qwen output.text/output.choices and MiniMax messages[].text response formats retained as fallbacks.

Tests

• Python: 344 passed ✅
• R: 237 passed ✅
• ruff: clean ✅

LGTM — ready to merge.