feat(skills): add verify-backwave verification skill

[pdevito3]

Summary

This change adds the project skill verify-backwave at .agents/skills/verify-backwave/. The .claude/skills symlink points to .agents/skills.

• SKILL.md has these sections: Launch, Doctor, Drive, Evidence, Cleanup, Helpers, and Gotchas.
• scripts/bwv is the run helper, with the commands launch, doctor, url, api, wait-job, browser, shot, sql, cleanup, and runs.
• features/ holds the feature map with five dashboard features: jobs and job detail, failures and requeue, queue pause and resume, workflows, and the tag filter.
• .gitignore adds .verify-evidence/.

The primary surface is the dashboard on Sample.Api with real jobs. The skill names the secondary surfaces: the Sample HTTP API and Swagger, the MCP server at /backwave-mcp, the BackWave.Demo host, Recurring Schedules, Observers, and Executing now. These surfaces have no recipe yet.

Isolation and cleanup

• Each run has its own id, a free port, a scratch directory, and a SQLite file. With --store postgres, the run starts its own labeled Postgres container.
• Each run has its own Chrome session and its own evidence directory, .verify-evidence/<run-id>/.
• bwv cleanup stops only the PID, the browser session, and the container that the run recorded. It never stops a process by name. It keeps the evidence and copies the final SQLite file into db/.

End-to-end proof

Run 20260927-154937-57901 (SQLite, port 55831) drove the queue pause and resume feature through real clicks in the dashboard:

1. bwv launch built the sample and printed READY. bwv doctor showed ok on every line.
2. A warmup job on critical reached Succeeded. The Queues page then showed critical and limited as Claiming (ui/pause-before-2.png).
3. A click on Pause in the critical row redirected to the Overview. The Overview showed PAUSED QUEUES 1 critical (ui/pause-overview.png).
4. A job enqueued on the paused queue stayed at state 0 (Scheduled), attempt 0, after 6 s (api/pause-stalled-job.json).
5. A click on Resume redirected to the Overview with PAUSED QUEUES 0 (ui/resume-overview.png). The stalled job reached Succeeded 1 s later.
6. db/pause-audit.txt has three rows: startup 5 limited, sample-operator 3 critical (Pause), and sample-operator 4 critical (Resume).
7. bwv cleanup stopped the browser session and app PID 58101, and removed the scratch directory. bwv runs shows the run as cleaned with app-live=no. The evidence directory still holds 27 files.

The proof found two faults in the first draft of the skill. Both are now corrected in the skill:

• A new run lists only limited on the Queues page. A queue shows only after it holds a job. The recipe now adds a job first.
• bwv shot takes a snapshot, so it makes the earlier uids stale. A click on a stale uid failed with STALE_REF, and the page did not change. The Drive rules now say to take the snapshot and click at once.

A second run, 20260927-155247-77205, verified the other recipe claims in the browser:

• The workflow list rows and the failing workflow graph. charge-payment is Dead-Lettered, and its dependents are Cancelled.
• A member's job page has no link back to its workflow.
• A requeue adds a Scheduled attempt-0 row and 3 new attempts, for 14 RECORDED rows.
• The tag suggestions go tenant ›, then acme, then ?tk=tenant&tv=acme.

That run was cleaned up the same way.

Out of scope

This change has no product code changes.