docs: improve repository navigation and skill discovery.

[pjgao]

Summary

• Add concise bilingual front doors and a curated documentation hub.
• Add presentation metadata for all 21 canonical skills without changing routing contracts.
• Generate role/domain skill views from the catalog and document repository evidence flow.
• Add one canonical maintainer workflow plus bounded navigation audit and usability evidence.

Before And After

Before, readers reconstructed ownership across root README, AGENTS, scripts, flat skill directories, and PR12 history. After, the root task table selects a primary entry in one page, docs/README.md separates current guidance from design history, skills/README.md groups capabilities by role/domain, and the maintainer guide provides one eight-step path.

Generated Artifacts

scripts/render_skill_docs.py generates skills/README.md and docs/architecture/skill-map.md deterministically. --check detects drift without rewriting files.

Compatibility And Scope

• 21/21 canonical skills and 19/19 baseline names preserved.
• Canonical IDs, directories, SKILL.md routing descriptions, and invocation policies unchanged.
• No nested physical skill categories.
• xLLM on Ascend remains the complete path; vLLM-Ascend/SGLang remain experimental adapter or artifact-analysis scopes.
• Presentation metadata is human-facing and does not make the catalog a runtime router.

Validation

• Catalog validation: PASS.
• Generated-doc check: PASS.
• pytest: 172 passed in 36.13s.
• git diff --check: PASS.
• Skill refresh: 21/21 installed, 0 broken links, 0 orphan SKILL.md.

Bounded Dogfood And Limits

Three isolated readers tested new-user optimization, operator discovery, and skill maintenance. After two navigation fixes, blocking dead ends were 0. This is a bounded qualitative sample, not broad usability or model-routing accuracy. Real campaigns still require detailed skills and a pinned spec; some operational skills are Chinese; historical examples outside PR14 navigation scope remain follow-up work.