Fix WMQ examples (78-84)

[mp-orkes]

Continuing the WMQ example pass from #492. Tested each against a local OSS server with WMQ enabled.

Bugs

• 78 — crashed on worker spawn (no if __name__ == "__main__" guard), and tempfile.mkdtemp() at module level gave every worker its own IPC dir, so approval requests never reached the main process. Dir now crosses via APPROVAL_WORKFLOW_IPC_DIR.
• 79 — dropped the last paragraph: the barrier waited on the Researcher forwarding, then stopped both agents while the Writer was still mid-turn. Now waits on publish sentinels.
• 80, 81, 82 — docstrings named anthropic/claude-sonnet-4-20250514, which 404s. On that model 82 never replies; it works on claude-sonnet-5.

Docstrings — corrected claims that no longer match the code: the wrong tool named as 78's approval gate, a closure in 79 that #477 removed, a "no time.sleep()" claim in 82_fan_out_fan_in contradicted by the code, updateVariables named as 84's stop mechanism (it's POST /agent/{id}/stop), undocumented /disconnect and /tools in 81, and a missing WMQ requirement throughout.

Testing

78: 5 tasks through 3 approval gates, COMPLETED. 79: 3/3 paragraphs across 4 runs (was consistently 2/3). 82_fan_out_fan_in: 9/9 answers, all 5 workflows COMPLETED. 83: same domain across runtime restart, both tasks ran. 84: _stop_requested: true, COMPLETED. 82: replies correctly on claude-sonnet-5.

79 and 82_fan_out_fan_in still fail roughly 1 run in 3 on #493 (handle.stop() leaving _stop_requested false) — unrelated to these fixes.

🤖 Generated with Claude Code

[mp-orkes]

Reviewed and manually tested all the examples with Claude’s assistance. The coding agent examples are interactive and were validated manually.