Add scoped agent write-intent actions

[cbusillo]

Problem

Agents sometimes need to request useful outcomes such as preview refresh, Every Code rerun, feedback acknowledgment, dry-run promotion, settings apply, or cleanup without receiving broad Launchplane write authority.

Scope

• Define write-intent action types and safety levels: safe_write, mutation, destructive, prod, and secret_backed.
• Prefer dry-run-first flows for actions that mutate runtime or configuration.
• Route runtime/product mutations through Launchplane while keeping GitHub as code/PR/comment/check truth.
• Require explicit elevated authorization for destructive, prod, and secret-backed actions.
• Record audit/provenance for agent-initiated intents.

Acceptance Criteria

• Agents can request safe scoped intents without a generic write token.
• Preview, rerun, dry-run, apply, cleanup, and promotion-dispatch candidates are individually authorized.
• Responses return status/evidence and source links, not reusable credentials.
• Tests cover allowed and denied intents across subject types and safety levels.

Relationships

• Subissue of Make Launchplane a read-only agent context provider #379.
• Depends on agent authn/authz and redaction/provenance contracts.