[hsliuustc0106]

Purpose

The README's quickstart starts only the frontend and leaves users to assemble a worker recipe. Add one complete Linux CPU LAYA walkthrough, a matching Chinese entry page, pinned tested dependencies, readiness checks and a real refund-decision demo. Put the walkthrough, demo and support matrix at the README entry point, and reconcile the matrix with the merged Python MPS worker and CLM stub-encoder recipe.

Refs #86 and #83. This PR remains draft for an independent contributor's first-use reproduction; it does not close #86 or claim that agent testing satisfies that acceptance item. No serving/model implementation changes or launcher are added.

Test Plan

System1-Omni Version / Commit: f594d7dfc4c2bef812e23f7ed73573be9625b287 (fetched upstream main; runtime/model code unchanged).

Reproduce the documented CPU worker/frontend path, check all three question types through direct and frontend requests, resolve CPU requirements, build the documentation site, check generated links/anchors, and compare the English/Chinese shell and response examples.

Test Result

• cargo build -p omni-jev --release --locked: passed.
• Real laya-serve CPU worker plus frontend: worker/frontend health and the documented refund request returned 200, both with the reused environment/checkpoint and with a fresh environment and isolated empty Hub model cache. The fresh run used current model revision 7b928d828b7b0e022f929d9bd2e44165aa270148 and produced the same outputs.
• python3.12 recipe/compare_with_backend.py --model english --backend http://127.0.0.1:8000 --frontend http://127.0.0.1:8086: all five checks passed (health, choice, score, noul and combined), with equal body bytes.
• CPU requirement resolver and existing-environment dry run: passed. The documented python3.12 -m venv .venv, CPU PyTorch installation and pip install -r recipe/laya/requirements-cpu.txt also passed in a fresh task-local environment; pip check found no broken requirements.
• mkdocs build --strict with the versions in docs/requirements.txt: passed.
• English/Chinese shell and JSON blocks agree; shell blocks parse with bash -n; local Markdown targets and rendered entry-page links/anchors resolve. git diff --check: passed.

GitHub CI passed the Docs build, benchmark-harness tests and Rust job (format, Clippy, frontend API tests, workspace tests and release build). The full workspace checks were not repeated locally for this documentation-only change. GPU execution is inapplicable. No new CPU performance or accuracy measurements are claimed. The existing scoped H200 result is linked separately from this CPU walkthrough.

Demo / evidence

See recipe/laya/validation.md, first-decision.json and first-decision-comparison.txt for the real request, complete response, comparison stdout, host/toolchain/package versions, model revision and preparation boundaries. The cached demo uses the English checkpoint at 55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851, Python 3.12.13 and CPU-only PyTorch 2.8.0. The host is Ubuntu 22.04.5 / Xeon Platinum 8480C; no GPU work was performed. Frontend port 8086 avoided an existing service on 8080. The fresh-run package capture is in recipe/laya/fresh-environment.txt; its five comparison outputs matched the cached transcript. Independent first-use reproduction remains pending. Video and new A/B performance measurements are N/A for this documentation change.

Self-review

Agent self-review used the repository's precheck-pr skill: base f594d7dfc4c2bef812e23f7ed73573be9625b287, head 2c3c595ac1918ee67df8c927088e00217701916c, clean working tree. The full documentation diff, technical scope, source support, examples and rendered links were checked. The contributor checklist remains for the contributor; this PR stays draft pending independent first-use evidence.

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.