[LauraGPT]

Summary

Add one dated bilingual article, hotwords-package-boundaries, explaining WebSocket hotword encoding, the standalone Torch/ONNX package boundary, and regression limits. Preserve all existing featured navigation; expose the new article in the archive and explanations section.

The archive browser contract now checks the catalogue count and every unique article URL, rather than a stale literal count. No visual threshold was relaxed.

Validation

• 85 focused Python tests pass. Two broader private runs reached their time bounds and are NOT claimed as full-suite passes. New-head and merged-main CI must each run all 543 site Python, 325 documentation, 46 legacy-link and 154 browser tests before deployment.
• Two deterministic builds: 275 identical files, 208 HTML pages validated. Desktop/mobile Chinese/English article checks pass in four cases; screenshots reviewed.
• Full private browser run: 150 passed, four failed solely on 37-versus-38 archive count. After the catalogue-based assertion change, all four affected cases pass; this is combined coverage, not a fresh full 154-case run. The original failures and a no-tests-selected runner error are retained.
• Static-only deployment helper: 13 passing tests. Publication uses an immutable directory and atomic symlink switch, without nginx/model-service reload, restart, configuration or log changes.
• Torch 0.1.7 must have accepted GitHub and PyPI bytes before publishing this article. Runtime evidence is synthetic contract coverage, not recognition quality, benchmark, production inference or adoption evidence.