workflows:work: Browser testing happens too late - tasks marked complete without visual verification

[zbeyens]

Problem

Claude marks tasks as "completed" without visual verification. Browser testing only occurs in:

• Phase 3: Quality Check (optional reviewer agents)
• Phase 4: Ship It (screenshots for PR)
• Later via /workflows:review

By then, multiple tasks may be "done" but actually broken. Claude is "blind" during implementation and guesses completion.

Current flow

Task loop → mark complete → next task → ... → Quality Check → Ship It
                ↑                                      ↑
            (blind guess)                    (first time seeing UI)

Proposed

Add browser verification IN the task loop for UI tasks:

Task loop:
  - Implement
  - Run tests
  - IF UI task: agent-browser verify → fix if broken
  - THEN mark complete

Suggested Change

In Phase 2 "Task Execution Loop", after "Run tests after changes", add:

- For UI tasks: run agent-browser to verify visually before marking complete
- If visual issues found, fix immediately (don't mark complete yet)

This catches issues per-task instead of accumulating blind failures until Phase 4.