[Menci]

Empty Responses reasoning summaries and Gemini thought parts could enter input IR and become { type: "thinking", thinking: "", signature: "" } on Messages targets. The reported Responses request contains 28 such empty summaries.

Filter empty reasoning before constructing input IR. The Responses reader removes empty text parts from summary and content and omits reasoning when both channels are empty. The Gemini reader skips empty thought text. The Messages reader skips empty thinking before segmenting visible text, so an empty block cannot split adjacent text. Chat Completions reasoning extensions already stay outside input IR; a regression test covers their empty forms.

Nonempty text and Gemini reasoning positions remain intact. Authenticated native replay continues to preserve signed thinking blocks whose text is empty.

Test Plan

• Regression tests reproduce empty Responses reasoning entering IR and Messages output, empty Gemini thoughts entering IR, and empty Messages thinking splitting adjacent visible text before the fix.
• Ten related test files pass (275 tests), covering all four source readers, Responses native replay, Responses-to-Messages, Responses-to-Chat, Gemini-to-Messages, Messages-to-Responses, and protocol round trips.
• Translate typecheck and ESLint on all changed files pass.
• Reading the captured request's 80 reasoning items produces zero input-IR reasoning items; 28 have empty-text summaries and the remaining 52 have no readable text.
• Authenticated native Messages replay preserves an empty thinking block and its signature.
• All 21 Verify CI checks pass for head 78c262d95689c9b226e5d36ef441f53c4812cee7.