fix(moderation): narrow the off-scope criterion to curb false positives

[albanm]

Rework the moderation gate's system prompt so its off-scope criterion
targets genuine misuse rather than any "heavy task."

What changed:

• The old rule blocked "a heavy task clearly unrelated to a data platform
  (e.g. generate a large program)", which over-fired on legitimate
  detailed/technical work — notably sub-agent task delegations.
• The off-scope criterion now targets using the assistant as a free
  general-purpose tool unrelated to the platform's data (general chatbot
  use, essays, substantial general-purpose software).
• Data exploration/analysis/visualization, working with dataset/file
  content, small scripts that consume the platform's data or API, and
  delegated sub-agent tasks stay in scope even when technical. Writing
  general-purpose code stays out; the ambiguous middle resolves to allow.
• Updates the unit test (asserts the in/out-of-scope wording) and the
  architecture doc.

Why: false-positive blocks were hitting sub-agents running targeted
technical/data tasks; the gate was too strict on legitimate detailed work.

Regression risks:

• The prompt is shared by every moderated org and the summary endpoint, so
  this widens the allow-set platform-wide, not just for sub-agent traffic.
• The classifier sees only message text and can't verify sub-agent
  provenance (modelRole is attacker-controlled, so it intentionally doesn't
  relax the policy). The small-script-vs-general-software line is a
  text-only judgment — expect some leakage in both directions.
• Mock-based API/e2e tests ignore the prompt, so the behavioral change is
  not covered by CI; only the prompt wording is asserted.