fix(mcp): persist and re-attach Gemini thoughtSignature on the direct Claude<->Gemini path

[Sam280903]

Summary

• The direct Claude<->Gemini translator (claude-to-gemini.ts / gemini-to-claude.ts) never persisted the thoughtSignature Gemini returns on functionCall parts, and never re-attached one on the next turn.
• Gemini 3+/2.5 strictly reject a native functionCall part with no signature (400). This surfaces in practice whenever a fallback combo lands on a Gemini model mid-conversation — the prior tool_use never went through Gemini, so no real signature was ever stored for it.
• gemini-to-claude.ts now stores the signature (keyed by tool_use id + connection namespace) when Gemini's response carries one, mirroring the existing gemini-to-openai.ts hub-path behavior.
• claude-to-gemini.ts now resolves a stored signature for historical tool_use blocks; when none exists and the target model requires one, the tool_use/tool_result pair is downgraded to inert text instead of being sent as a signature-less native part — matching the "context" fallback the OpenAI hub path already uses for the same situation (fix(gemini): resolve truncation/suppression of false positive textual tool call markers in backticks #3358), rather than the removed fake-signature injection this file's dead code comment referenced.

Test plan

• node --import tsx/esm --test tests/unit/translator-claude-to-gemini.test.ts — updated the two pre-existing tests that relied on the old (buggy) signature-less passthrough to prime a real stored signature, added a regression test reproducing the exact 400 (signature-less tool_use/tool_result pair is downgraded to inert text on a thinking-capable model), and a test locking in that older non-thinking models are unaffected.
• node --import tsx/esm --test tests/unit/translator-resp-gemini-to-claude.test.ts — added a test asserting the signature from a standalone thoughtSignature part is persisted and retrievable via getGeminiThoughtSignature.
• node --import tsx/esm --test tests/unit/vertex-functioncall-id-3440.test.ts — updated the one affected pre-existing test (primed a signature; unrelated to what it actually tests).
• Full run across all touched/adjacent test files: 51/51 passing.
• npm run lint clean on all touched files (test-file any usage stayed within the frozen per-file budget).
• npx tsc --noEmit clean on all touched files.

[diegosouzapw]

Review: PR #9448 — fix(mcp): persist Gemini thoughtSignature on direct Claude<->Gemini path

Verdict: merge-ready (5/5 stars)

Correctly persists and re-attaches Gemini thoughtSignature. Matches existing hub-path behavior. Tests included.

Ready to merge after retarget/rebase to release/v3.8.50 tip.

[diegosouzapw]

Merged into release/v3.8.50 — thank you @Sam280903 for the contribution! It was validated on a combined merge-train (static gates + affected tests + vitest) together with 30 sibling PRs before landing. The release had independently landed a superset of this fix meanwhile, so what shipped from your branch is the extra regression test covering the standalone thoughtSignature part — valuable coverage the suite did not have.