fix(openai-compat): drop top_k for active Claude thinking

[smartershining]

Summary

Remove top_k from OpenAI-compatible payloads when a Claude request has active thinking.

Claude clients may send top_k together with extended/adaptive thinking. After Claude-to-OpenAI translation and payload overrides, that combination can reach a Claude-backed compatibility upstream and be rejected with -4003 top_k must be unset when thinking is enabled/adaptive.

This change runs after payload overrides and:

• applies only to Claude-source requests;
• treats thinking.type=enabled|adaptive|auto and non-none effort fields as active thinking;
• removes only top_k;
• preserves top_k for non-Claude, no-thinking, and explicit-none requests;
• keeps sibling sampling fields unchanged.

Both streaming and non-streaming execution paths are covered.

Tests

• go test -count=1 -run 'TestNormalizeOpenAICompatClaudeSampling|TestOpenAICompatExecutorUsesCompatibleClaudeTranslation|TestNormalizeClaudeSamplingForUpstream' ./internal/runtime/executor
• go build -buildvcs=false -trimpath -o /tmp/cpa-pr2-build ./cmd/server

[github-actions]

This pull request targeted main.

The base branch has been automatically changed to dev.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 23546faf49

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Respect final thinking overrides before dropping top_k

When a Claude request initially enables thinking but a model suffix such as (none) or a payload override disables it, ApplyRequestThinking and the override stage produce a final non-thinking request, yet this switch still marks thinking active from the pre-override source and deletes top_k. This silently changes a valid sampling configuration even though the upstream-invalid combination no longer exists; determine activity from the finalized canonical/transformed body so the established suffix and override precedence is preserved.

AGENTS.md reference: AGENTS.md:L29-L29

Useful? React with 👍 / 👎.