fix: signature + correct field for streaming reasoning thinking blocks

[angelz07]

What

Fix empty Claude Code responses from reasoning models exposed via the OpenAI-style reasoning_content field (LM Studio local models, OpenAI o1/o3).

Why

When a backend streams reasoning via delta.reasoning_content, the proxy produced a thinking block that Claude Code silently discarded, leaving the user with an empty response even though the model returned valid text. Two defects in streamOpenAIToClaude:

1. The thinking_delta used the field text instead of thinking, so the reasoning payload was malformed. The OpenRouter reasoning_details path already used thinking correctly.
2. No signature_delta was ever emitted, and the thinking block was closed only at stream end — after the text block. Claude Code discards thinking blocks that arrive without a signature (see the pitfall documented in CLAUDE.md), so the whole assistant turn was dropped.

How

Add a reusable closeThinkingBlock() helper that emits a signature_delta followed by content_block_stop, and call it before any text/tool block starts and once more at stream end. This finalizes the thinking block in the order Claude Code expects:

content_block_start(thinking) -> thinking_delta* -> signature_delta -> content_block_stop -> content_block_start(text) -> ...

The signature value is a fixed opaque string; it is never validated against Anthropic for local/OpenAI-compatible providers, and Claude Code only requires its presence to treat the block as valid thinking.

Testing

• go test ./internal/server/... passes.
• Verified end-to-end against a local LM Studio (Qwen3 reasoning model): raw Claude SSE now shows thinking -> signature_delta -> stop -> text ordering with the text content preserved, where before the thinking block had no signature and the wrong delta field.

🤖 Generated with Claude Code