fix: harden web reads and complete P1 cleanup

[Panniantong]

Summary

• reject non-public and ambiguous Web targets before Jina Reader, including legacy IPv4 spellings; cap responses at 5 MiB
• detect high-confidence Jina/Cloudflare challenge pages and fail explicitly without unsafe direct fallback
• make packaged skills discoverable by OpenCode and conform both frontmatter variants to its schema
• remove retired Douyin and Weibo claims from Japanese and Korean docs

Verification

• Python 3.10: 475 tests passed
• Python 3.12: 475 tests passed
• Python 3.13: local wheel built, installed, version checked, and OpenCode discovery smoke passed
• changed-file Ruff checks passed
• git diff check passed
• independent read-only review found no blocking P0/P1 issues

Context

This safely reimplements the still-useful parts behind #476, #480, and #483 from current main. It intentionally does not add direct target-site fallback because that proposal would expose the user network and introduce redirect/SSRF and unbounded-response risk. The challenge fixture is based on #428.