Fix incorrect/incomplete command examples in the agent-workflow guide

[johnjosephhorton]

Severity: medium — following the guide verbatim errors.

• Stage 8 edsl-export --target twin-probability-job example omits --approved-plan/--allow-unapproved, so run verbatim it fails approval_required — contradicting the guardrail two sections up. Add the approval flag to the example.
• Stage 3 raw add example omits the required --title (argparse marks it required), so the example fails.
• maxOutputTokens: "Prefer maxOutputTokens=8192" doesn't note it's Google-specific. Passing it globally makes OpenAI warn Unknown parameter(s) for model 'gpt-5.5': maxOutputTokens (OpenAI wants max_tokens). Show the per-model form --model-param google:gemini-2.5-pro:maxOutputTokens=8192.

---

Found while building a digital-twin validation for a 33-respondent SMB survey (smb.csv) end-to-end via zwill guide (survey → import → commit → one-shot marginals → twin-probability job → twin-validate → report). File/line refs are from zwill/ at that checkout.