feat: package SkillOS demo-paper evaluation flow

[dydpop]

Summary

• Add a repository-level Windows quick start and one-click demo launcher path: install deps, run START_SKILLOS_DEMO.bat, enter local DeepSeek API URL/model/key, then SkillOS opens in the browser.
• Add demo restore/readiness/evaluation scripts for the five-input SkillOS flow: trajectory, document, api_doc, script, and past_skills.
• Integrate selected PR Feat/lbs repo version #11 Git/version ideas on the current branch: business-readable diffs, editable version fields, Version Lab, and re-verification when implementation/interface changes.
• Improve the frontend graph into a more demo-friendly Nebula/Readable/Debug view with node/edge/label controls.
• Add PR-ready docs for group operation, update summary, upload scope, and SkillsBench five-task analysis.

SkillsBench evidence

Selected 5-task P0 sparse subset:

• Oracle: 5/5
• No-skill baseline: 2/5
• SkillOS generated skills: 3/5 using clearly labeled clean combined evidence

The clearest gain is dialogue-parser, from 0.667 to 1.0. The two clean failures are sales-pivot-analysis and software-dependency-audit, which now define the next repair-loop targets: exact Excel pivot metadata and exact security CSV normalization.

Verification

• Secret-pattern scan over staged/source docs: no real sk-... key pattern matches.
• PowerShell parse check for skillos-one-click-launcher/scripts/Start-SkillOSDemo.ps1: passed.
• python -m compileall for updated backend/eval scripts: passed.
• Targeted backend tests: 45 passed.
• Frontend npm run lint: passed.
• Frontend npm run build: passed, with only the existing Vite large chunk warning.
• git diff --check: only LF/CRLF warnings on Windows.

PR scope notes

This PR intentionally includes reproducible code, tests, lightweight demo fixtures, scripts, and summary reports. It intentionally does not include raw sb-runs benchmark logs, Docker installers/images/cache, runtime logs, local launcher config, or API keys.

Limitations

• The SkillOS generated SkillsBench 3/5 result is clean combined evidence, not one uninterrupted all5 generated-skill run.
• Ctx2Skill support is a demo-paper Ctx2Skill-lite implementation, not full large-scale self-play.
• The next research/engineering step is to use the two benchmark failures as Skill repair-loop demonstrations.