Add an optional Agent Assistant policy bridge for confirmed turn actions

[khaliqgant]

Summary

@agentworkforce/turn-kit provides runConfirmedTurnAction() to ensure success text is constructed only after a provider mutation returns a confirmed receipt. It does not yet provide reusable risk classification, approval gating, or audit records.

Add an optional bridge to @agent-assistant/policy so agents can apply consistent mutation policy without coupling the core turn lifecycle to Agent Assistant.

Depends on the turn-kit foundation in #299.

Scope

• Add an optional turn-kit policy entry point, likely @agentworkforce/turn-kit/policy.
• Map a requested turn action into Agent Assistant action classification and gating.
• Support allow, deny, and explicit-approval outcomes before the mutation runs.
• Compose with runConfirmedTurnAction() so provider receipts remain the source of truth for success.
• Emit a normalized audit result for requested, denied, approved, attempted, confirmed, and unconfirmed outcomes.
• Keep @agent-assistant/policy optional; the base turn-kit package must work without it.
• Avoid embedding provider-specific GitHub, Linear, Telegram, or Relayfile logic.

Acceptance criteria

• Denied or unapproved actions never invoke the provider mutation.
• Approved actions still cannot produce success text until the provider receipt is confirmed.
• Unconfirmed provider receipts are audited as failures and retain the current fail-closed behavior.
• Policy metadata can include conversation, actor, action class, and provider reference without storing sensitive message content by default.
• Tests cover low-risk allow, high-risk approval, denial, provider failure, and unconfirmed receipt paths.
• Documentation includes a complete mutation example and explains the ownership boundary between policy and provider confirmation.