fix(creation-helper): stabilize Gemini tool turns

[rppavan]

Preserve raw Gemini function-call parts so thought signatures survive tool-response turns.

Map raw Gemini function calls back to tool result names, guard against repeated successful tool calls in the Smart Creator loop, and treat tool-only assistant turns as valid UI output instead of showing a false failure banner.

Verified with npm run build, cargo test preserves_raw_function_call_parts_with_thought_signatures --lib --release, codesign verification, hdiutil verify, and mounted-DMG app signature verification.

[MegalithOfficial]

I think the gemini_content replay fix is directionally correct, but I see a regression in the new duplicate-tool suppression.

In src-tauri/src/creation_helper/agent/run.rs, the completed_tool_signatures guard stops the turn on a repeated successful (name, args) signature before we append any CreationToolResult for the current tool call. That means if Gemini re-emits the same function call on the next iteration, we now continue past dispatch, set terminal_hit = true, and then break the loop without ever answering the model’s tool call with a tool result. That breaks the tool-calling handshake in the exact retry case this PR is trying to stabilize.

I also think the dedupe is too broad for the fix scope. The replay issue is Gemini-specific, but this changes behavior for every provider and for all tools. Some creation-helper tools are intentionally idempotent or flow-oriented (show_preview, request_confirmation, set_*), so repeating the same call in one turn should be harmless; turning that into an early stop looks like a behavioral regression.

I'd keep the raw gemini_content preservation, but either remove the duplicate-call suppression or narrow it to a safer Gemini-specific path that still returns a valid tool result for repeated calls.