fix(claude): use canonical reasoning_content in non-stream response

[warelik]

Problem

When translating non-streaming Claude responses to OpenAI Chat Completions (ConvertClaudeResponseToOpenAINonStream), thinking blocks were written to the non-canonical field choices.0.message.reasoning. In contrast, the streaming converter in the same file (ConvertClaudeResponseToOpenAI) outputs reasoning deltas using the canonical choices.0.delta.reasoning_content.

Because downstream OpenAI clients and multi-turn request adapters expect choices.0.message.reasoning_content, non-streaming responses failed to preserve thinking tokens across subsequent turns.

Fix

• internal/translator/claude/openai/chat-completions/claude_openai_response.go:437: Updated ConvertClaudeResponseToOpenAINonStream to serialize reasoning content to choices.0.message.reasoning_content.

Tests

• TestConvertClaudeResponseToOpenAINonStream_ReasoningContent: Verifies reasoning_content is properly populated and legacy reasoning field is omitted.
• TestConvertClaudeResponseToOpenAINonStream_OmitsReasoningContentWhenAbsent: Verifies reasoning fields are omitted when responses lack thinking blocks.
• TestConvertClaudeResponseToOpenAI_StreamAndNonStreamParity: Verifies stream and non-stream parity for thinking content, visible text, and finish reason.
• TestConvertClaudeResponseToOpenAI_RedactedThinkingIgnored: Verifies redacted_thinking blocks are never leaked into reasoning_content in streaming or non-streaming mode.

Reverse bite-check

Reverting internal/translator/claude/openai/chat-completions/claude_openai_response.go to the pre-fix state reproduces the failure where reasoning_content is missing:

=== RUN   TestConvertClaudeResponseToOpenAINonStream_ReasoningContent
    claude_openai_response_test.go:226: expected choices.0.message.reasoning_content to exist, payload={"id":"msg_123","object":"chat.completion","created":1787286777,"model":"claude-opus-4-6","choices":[{"index":0,"message":{"role":"assistant","content":"Here is the solution.","reasoning":"Let me analyze the problem. Step 2 is clear."},"finish_reason":"stop"}],"usage":{"prompt_tokens":10,"completion_tokens":20,"total_tokens":30,"prompt_tokens_details":{"cached_tokens":0,"cached_creation_tokens":0}}}
--- FAIL: TestConvertClaudeResponseToOpenAINonStream_ReasoningContent (0.00s)
=== RUN   TestConvertClaudeResponseToOpenAI_StreamAndNonStreamParity
    claude_openai_response_test.go:317: parity mismatch for reasoning_content: nonStream="", stream="First thought. Second thought."
--- FAIL: TestConvertClaudeResponseToOpenAI_StreamAndNonStreamParity (0.00s)
FAIL
FAIL	github.com/router-for-me/CLIProxyAPI/v7/internal/translator/claude/openai/chat-completions	0.385s
FAIL

Verification

TMPDIR=/Users/warelik/.cache/gotmp GOCACHE=/Users/warelik/.cache/gocache go build ./...
TMPDIR=/Users/warelik/.cache/gotmp GOCACHE=/Users/warelik/.cache/gocache go vet ./internal/translator/claude/openai/chat-completions/...
gofmt -l internal/translator/claude/openai/chat-completions/claude_openai_response.go internal/translator/claude/openai/chat-completions/claude_openai_response_test.go
TMPDIR=/Users/warelik/.cache/gotmp GOCACHE=/Users/warelik/.cache/gocache go test -v -run "TestConvertClaudeResponseToOpenAINonStream_ReasoningContent|TestConvertClaudeResponseToOpenAI_StreamAndNonStreamParity|TestConvertClaudeResponseToOpenAI_RedactedThinkingIgnored" ./internal/translator/claude/openai/chat-completions/...

Output:

=== RUN   TestConvertClaudeResponseToOpenAINonStream_ReasoningContent
--- PASS: TestConvertClaudeResponseToOpenAINonStream_ReasoningContent (0.00s)
=== RUN   TestConvertClaudeResponseToOpenAI_StreamAndNonStreamParity
--- PASS: TestConvertClaudeResponseToOpenAI_StreamAndNonStreamParity (0.00s)
=== RUN   TestConvertClaudeResponseToOpenAI_RedactedThinkingIgnored
--- PASS: TestConvertClaudeResponseToOpenAI_RedactedThinkingIgnored (0.00s)
PASS
ok  	github.com/router-for-me/CLIProxyAPI/v7/internal/translator/claude/openai/chat-completions	0.388s

Tooling note

The jbcontext semantic-search CLI is installed but non-functional in this environment
(jbcontext search fails with the OS keychain is not accessible). The equivalent
review passes were performed with the repository's own tooling and manual inspection
instead; this is disclosed for transparency about how the change was reviewed.