fix(translate): route OpenAI-compat reasoning to Anthropic thinking blocks

[steventohme]

Problem

Qwen models routed through OpenRouter (and DeepSeek native) emit reasoning traces in delta.reasoning / delta.reasoning_content. The AnthropicSSETranslator only read delta.content, so reasoning either silently dropped or — when the model embedded its planning narration inline — bled into the visible text channel. Observed symptom: Claude Code clients saw repeated "Let me check..." narration mixed into the answer instead of a clean response.

Fix

internal/translate/stream.go

• New state field thinkingOpen plus emitContentBlockStartThinking / emitContentBlockDeltaThinking helpers.
• emitDelta reads reasoning_content (DeepSeek/Qwen native) then reasoning (OpenRouter normalized) and emits Anthropic thinking content blocks. The block-open state machine closes thinking before opening text / tool_use (and vice versa), with finishStream closing any still-open thinking block.

internal/translate/translate.go

• writeAnthropicContentFromOpenAI emits a thinking block before the text block when the buffered (non-streaming) response carries reasoning / reasoning_content.

Tests

Added in translate_test.go:

• TestAnthropicSSETranslator_StreamingReasoningEmitsThinkingBlock — reasoning deltas followed by content produce a thinking block then a text block in that order; reasoning prose is asserted to NOT appear in the text channel.
• TestAnthropicSSETranslator_StreamingReasoningContentField — reasoning_content (DeepSeek convention) maps to the same thinking block.
• TestAnthropicSSETranslator_NonStreamingReasoningEmitsThinkingBlock — buffered responses emit thinking block before text block.

go test ./internal/translate/... and go test ./internal/proxy/... both pass.

Test plan

• CI green
• Manual check: route a Qwen3 (or DeepSeek-V4-Pro) turn via the router and confirm Claude Code renders reasoning in its thinking pane rather than as inline assistant text.