fix(antigravity): preserve valid tool-call history

[victorwads]

Summary

Fixes two Antigravity/Gemini tool-history serialization failures:

• Normalizes every functionResponse.response value to a JSON object, as required by the Gemini API. Arrays, primitives, and null are wrapped as { result: value }; non-JSON text remains { output: text }.
• Preserves real replayed thought signatures and supplies Gemini's documented skip_thought_signature_validator sentinel for the first unsigned function call in each Gemini 3 tool-call step. This permits users to switch an existing conversation from another provider (for example GPT/Codex) to Gemini 3 without a 400 validation failure.

Context

When serialized tool results have a top-level array, Antigravity rejects the request with an error such as:

Invalid JSON payload received: Proto field is not repeating, cannot start list.

When a prior provider's tool call is replayed without a Gemini thought signature, Gemini 3 rejects it with:

Function call is missing a thought_signature in functionCall parts.

The fallback sentinel is the documented approach for manually constructed or cross-model histories that do not carry Gemini signatures.

Tests

• Covers object, array, null, and numeric tool-result payloads.
• Covers parallel and sequential foreign tool-call steps replayed by Gemini 3.
• Verifies that Gemini 2.5 does not receive the Gemini 3 fallback.

Closes #97

References:

• https://ai.google.dev/api/generate-content
• https://ai.google.dev/gemini-api/docs/generate-content/thought-signatures

[victorwads]

@V1ki i'm using this branch for some days, so, i think maybe it should be merged. tested A LOT on several antigravity sessions.

Most tests changing from Codex to Antigravity to Codex to Antigravity mid session.
It Works!

[V1ki]

感谢 @victorwads 的贡献，也感谢在 #97 里把两个问题梳理清楚！

我先合并了 #100（它修的是同一个 functionResponse.response 非对象的问题，测试覆盖更全），然后把这个分支 rebase 到了最新 main（4278b1d）：

• toolResultValue 采用 fix(antigravity): keep functionResponse.response a protobuf Struct #100 的 { output: value } 封装，与现有非 JSON 路径保持一致，同时保留了你把返回类型收紧为 Record<string, unknown> 的改动；
• thought signature 部分原样保留。我对照 Google 文档核实过，skip_thought_signature_validator 是官方占位签名，"每个 step 只校验第一个 functionCall、仅 Gemini 3 严格校验"的规则也和实现一致；
• 原有测试里 result 键的断言改成了 output，其余不变。干净构建下 501 个测试全部通过。

方便的话请再看一眼 rebase 后的分支，没问题我就合并。

---

Thanks @victorwads for the contribution, and for laying out both failures so clearly in #97!

I merged #100 first (it fixes the same non-object functionResponse.response issue with broader test coverage), then rebased this branch onto the latest main (4278b1d):

• toolResultValue now uses fix(antigravity): keep functionResponse.response a protobuf Struct #100's { output: value } envelope, consistent with the existing non-JSON path, while keeping your tightened Record<string, unknown> return type;
• the thought-signature part is unchanged. I verified against Google's docs that skip_thought_signature_validator is the official placeholder and that the "first functionCall per step, strict on Gemini 3 only" rule matches the implementation;
• the result key assertions in the existing test were switched to output; nothing else changed. All 501 tests pass on a clean build.

Please take a quick look at the rebased branch when you have a moment. If it looks good to you, I'll merge it.

[victorwads]

Thanks