[sunrioa]

Summary

The request encoder's async state can make the surrounding futures large enough to overflow a worker stack. The encoder now owns its pinned heap allocation: emit returns a concrete, statically dispatched future, and the counting, in-memory build and backed encoding paths await it directly. All paths continue to use the existing codec.

Verification

• 29 maka-model unit tests and 8 integration tests passed on macOS arm64.
• Target-crate Clippy with warnings denied, formatting and diff checks passed.
• An isolated dev-build probe completed text and tool/media request preparation plus full wire replay with 512 KiB and 2 MiB thread stacks. Wire bytes, identities and SHA-256 digests matched the original PR, and allocation counts were unchanged. The public prepare_source future remained 2,184 bytes on Rust 1.99.1.
• The original PR's invalid-image and compaction integration investigation is not independently rerun by this follow-up; the probe is a focused codec check.

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex assisted with investigation, implementation, verification, and drafting, including centralizing the heap allocation in emit.

Checklist

• Tests cover the change and fail without it
• Format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes
• No

简体中文

编码器的 async state 会增大外层 future，导致工作线程栈溢出。现在由 emit 自己统一在堆上分配 pinned future，计数、内存构建和 backed 编码三个入口直接 await，保留原编码逻辑和静态分发。

macOS arm64 上通过 29 项模型单元测试、8 项集成测试、目标 crate 的 Clippy、格式和 diff 检查。独立 dev probe 在 512 KiB 和 2 MiB 线程栈上完成文本及 tool/media 请求准备和完整 wire replay，输出字节、identity、SHA-256 与原 PR 一致，分配次数不变。此前 invalid-image 和 compaction 的整机联调场景未在这次封装调整中独立重跑。

OpenAI Codex 协助排查、实现、验证和正文起草。