docs: 添加了miloco推理管线onnx需要本地推理的说明

[Kylehub-Kyle]

• README (EN/ZH): update caution block to clarify local ONNX inference is required, not purely cloud-based

[github-actions]

👋 感谢提交 PR @Kylehub-Kyle！维护者会尽快 review。

提交前请确认：

• CI 全绿（test / lint / build）
• 改动聚焦单一主题，便于审阅
• 若改动了依赖（lockfile / pyproject.toml / package.json），需维护者评论 /allow-dependencies-change <当前 head SHA> 放行（之后再 push 需重新放行）

[github-actions]

PR #388: docs: 添加了 miloco 推理管线 onnx 需要本地推理的说明

作者: Kylehub-Kyle
范围: docs/add-cpu-requirements → main

修改方案

纯文档改动（README.md / README.zh.md 各 +3/-1）。把原来单段「成本 / 费用提示」CAUTION 块拆成两段：

• 第一段（API 费用提示）：保留原意——感知与 Agent 主要依赖云端大模型，会持续产生 API 费用，请在家庭面板「模型」页关注 token 用量。
• 第二段（本地 ONNX 推理必需，新增）：补充说明感知管线的人体检测 / 重识别模型经 ONNX Runtime 在本机运行，"并非全部依赖云端 / not purely cloud-based"，纠正读者可能产生的"纯云端"误解。

commit 线（5 条）：clarify local ONNX inference requirement → reorder caution blocks — API cost first, then local ONNX note → 三条 Update README（含对上轮 ci review 语法问题的修订）。此前含 CPU 门槛详表、维护者 @yangbaofu007 追问「具体设备实测」的内容已整体撤下，本轮 diff 只剩上述两段澄清。

复核记录（本轮）

• 上轮 ci 🔵 已修复：上一轮 review 指出英文侧改写手滑成病句 rely primarily use cloud large models。本轮 diff 已修正为 rely primarily **on** cloud large models（README.md:44），语法成立、并与中文「依赖」口径对齐 ✅
• 技术声明属实（C 跨层一致性）：README 断言「检测 + 重识别经 ONNX Runtime 本地运行、并非全部依赖云端」，与 origin/main knowledge/05-external-deps/sdk-onnxruntime.md 完全一致——该文档写明「Miloco 用 ONNX Runtime 在本地运行感知流水线中的多类模型……无需外部 API 调用」，且 det_4C.onnx（检测，必需）、human_body_reid_v2.onnx（ReID，必需）真实存在、由 make_session（perception/inference/ort_utils.py）建 CPU EP session ✅
• 两段无自相矛盾：第一段「云端大模型」指多模态 LLM（MiMo）推理，第二段「本地 ONNX」指检测 / ReID，指向感知管线不同环节，并存不冲突 ✅
• 文档声明对账（B 文档层）：PR body 唯一声明「update caution block (EN/ZH) to clarify local ONNX inference is required」在两个 README 里都兑现 ✅；5 条 commit message 与实际改动一致（含 reorder 后 API 费用在前的顺序、以及修语法的 Update README）✅
• 维护者遗留问题：@yangbaofu007 追问的「具体设备实测出处」所指详表已随内容撤下，问题随之消解 ✅

问题

本轮无 🔴 / 🟡 / 🔵 问题——上轮唯一的 🔵（英文语法病句）已在本轮 diff 修复。

结论

LGTM — 纯文档澄清。核心技术声明经 origin/main 知识库复核属实（检测 / ReID 为必需的本地 ONNX 模型），两段无冲突，PR body 与 diff 一致，5 条 commit message 对齐；上轮 ci 提的英文语法问题（rely primarily use）本轮已修为 rely primarily on，EN/ZH 口径一致；维护者遗留追问已随详表撤下而消解，无阻塞项。

---

_{由 review-pr skill v1.6 生成}

[yangbaofu007]

@Kylehub-Kyle 感谢提交PR，请关注下ci review 提的不一致问题。另外更新了具体的设备，有具体的实测么？

[ExWang]

Thanks for the PR! Spelling out the CPU / local-inference requirements is genuinely helpful — "why do I need a decent CPU if it's cloud-based?" trips up a lot of first-time deployers. Good direction 👍

One thing I'd like to align on before merging, plus two suggestions:

1. [Needs fix] The three docs disagree on the x86 minimum

• sdk-onnxruntime.md table: x86 Linux/Windows → minimum Celeron N5105 / recommended i3 10th-gen+
• but dev-guide.md says x86 i3 8th-gen+, and README (EN + ZH) says "x86 needs i3 8th-gen or newer"

These don't line up, so a reader can't tell whether an N5105 actually qualifies (the table says yes, README/dev-guide imply no). It looks like the x86 macOS row's i3-8th-gen got treated as the bar for all x86, dropping the x86 Linux/Windows row. Suggest making the table the single source of truth and having the README/dev-guide summaries either say "see table" or distinguish the two x86 platforms the same way the table does.

2. [Question] Are the x86 tiers measured or estimated?
I saw from the group chat that you've tested the Raspberry Pi tiers, so this is just about the x86 rows. The only value flagged as "empirically calibrated" in the diff is the ONNX thread count; the x86 CPU tiers (N5105 minimum, i3-10th-gen+ recommended, Atom/Z8350 ">200ms") don't cite a basis. If they're measured, could you note the test hardware? If they're estimates, I'd suggest softening "minimum/recommended" to "rough guidance" and noting they haven't been validated across all chips — so nobody buys an N5105 as the "minimum" and finds it can't keep up.

3. [Suggestion] Ordering of the CAUTION block
The original CAUTION contained only the cost note; this version puts the new "local ONNX inference required" paragraph first and pushes the cost note down to second. Order affects what the reader's eye lands on first. If there wasn't a specific reason, I'd suggest appending the new ONNX paragraph after the cost note (keeping cost first); if you do think local inference deserves top billing, that's fine too — just call out the trade-off in the description. Also, "not purely cloud-based" sitting right next to "cloud-based models … incur costs" reads a little at odds; might be worth a light polish.

(The "Hardware" bullet inserts CPU right after storage and keeps the existing order intact — that part's fine.)

Thanks again 🙏

[Kylehub-Kyle]

很感谢大佬能抽空审查，这是我第一次使用GitHub，仍有许多操作不明白，依赖群内机器人操作，很担心会因此给研发组带来额外的工作负担。真心感谢大佬的审核，真的学到很多。

[yangbaofu007]

@Kylehub-Kyle 感谢你的PR，caution block 是专门用来提醒用户注意的事项，这里是注意成本，onnx 本地推理不应该放到 caution block里面，不需要专门提醒，但可以写到其他地方，做个说明是可以的。