feat(cli): add agent-reach route for URL to channel/backend/command

[AbdulrahmanAmer]

Follow-up to #613, separate branch so you can take them independently.

While I was in the channels code for the doctor change I hit something else. When my agent has a URL in hand, there's no way for it to ask which platform owns it and what to run. That mapping only exists as prose in SKILL.md plus the reference docs, so the agent has to read the skill and work it out. doctor --json tells it what's healthy, never what handles this particular link.

So, agent-reach route <url>, with --json for the agent:

$ agent-reach route https://www.xiaohongshu.com/explore/abc
URL      : https://www.xiaohongshu.com/explore/abc
渠道     : xiaohongshu — 小红书笔记
后端顺序 : OpenCLI、xiaohongshu-mcp、xhs-cli (xiaohongshu-cli)
命令     :
  [OpenCLI] opencli xiaohongshu note "https://www.xiaohongshu.com/explore/abc" -f yaml
  [xhs-cli (xiaohongshu-cli)] xhs read "https://www.xiaohongshu.com/explore/abc"
参考文档 : agent_reach/skill/references/social.md
提示     : 运行 `agent-reach doctor --json` 查看当前实际生效的后端

The part I'd most want your opinion on is where the commands come from. I did not want this thing guessing CLI flags, so there's a per-channel url_commands map and I only filled in backends where your own references show the tool taking a URL or ID: twitter-cli tweet URL_OR_ID, opencli xiaohongshu note NOTE_URL, xhs read NOTE_ID_OR_URL, yt-dlp, Jina Reader, feedparser. Bilibili, Reddit and V2EX read by id or keyword, so they print no command at all and send the agent to the reference doc. Printing something that looks right but doesn't run seemed worse than printing nothing.

Rest of it:

• WebChannel.can_handle() is True for everything, so routing skips it while matching and only falls back to it. Otherwise where it sits in the registry would decide every answer.
• if a channel's can_handle throws, it gets skipped rather than breaking routing for everything else. Same thing doctor already does with check().
• no subprocess, no network, no writes, Config(read_only=True). It's meant to be cheap enough to call per URL. Two tests hold that line: one makes subprocess.run and Popen raise, one asserts the config dir never gets created.
• goes through the existing ordered_backends(), so a <channel>_backend override moves the command to the front too and route agrees with doctor.

Also gave every channel a reference field so the routing table has a machine-readable pointer to the doc the agent should read next. There's a test that opens each one, so it fails if a reference doc gets renamed.

Documented in both SKILL.md and SKILL_en.md under the environment check section, since that's where an agent is already looking.

571 existing tests still pass, 41 new ones in tests/test_route.py: routing for all 15 channels, the fallback, bad input, {url} substitution and ordering, the override, plus consistency checks that no url_commands key names a backend that isn't declared and every template actually contains the placeholder.