MANUAL RUN: try the dependency doctor's on-fail chain for real (#869) #873

Follows #869 (the on-fail chain from #861). The tests only use a fake tester and fixer, so these steps need a real repo, a real claude login and the page itself.

Setup: on the desktop app, take a test repo whose tests break on a dependency update. For example, pin "clsx": "~1.2.1" and add a test that fails on clsx 2, with a 1.3.0 release available to roll back to. On its repo page, turn tests on, set AI to claude, and build the chain 🔁 🤖 🔎 ↩️ 📌 🔔.

• Chips: tapping one toggles it and ‹ moves it earlier. Save and reload, and the order sticks. Nothing wraps or shifts at phone width.
• 🤖 and 🔎 on real claude: claude now runs with --output-format json. Check that its edits still come back and that the report shows a $ cost on the 🤖 and 🔎 steps when you hover their icons.
• 🔎 searches: claude's answer names a source it found, such as an issue, release note or migration guide.
• Dollar cap: set it to $0.05 and run again. 🔎 shows "out of dollars".
• ↩️ roll back: the update lands on the branch at the smaller version, and the report reads "rolled back, tests pass".
• 📌 hold back: the release goes into the "held back" box, and the next "Check now" doesn't propose it again.
• 🔔 alert: with only 🔔 on, a macOS notification shows up once, and not again on the next run.
• Hosted (Cloudflare, admins): redeploy so the container picks up the new sandbox/server.mjs. Then check that 🔎 works through the egress gate with no new hosts, and that the cost adds up.

File anything that breaks as its own issue and link it here.