Define agent classification scheme (Class 1–4) and pre-production approval gate

## Description

Establish a classification framework for agentic AI systems based on consequence and autonomy (Class 1: informational/read-only output, Class 2: reversible internal actions, Class 3: irreversible or high-consequence actions, Class 4: prohibited), and a mandatory approval step before any agent is classified and deployed to production.

## Source

Retningslinje for agentisk KI i Digdir v1.0 (15.06.2026), §5.1 "Klassifisering av agentiske løsninger" and Krav 1 "Klassifisering før idriftsetting": every agentic solution must be classified before going to production, and the classification must be recorded and approved before deployment.

## Business Goal

Without a shared classification scheme, teams have no consistent way to decide which governance controls (approval gates, kill-switch, DPIA, etc.) apply to a given agent — undermining every other control in this epic.

## Acceptance Criteria

- A documented classification scheme (Class 1–4) exists, matching the guideline's consequence/autonomy criteria, with worked examples.
- A classification field exists in the `.agent-card.yml` schema (see the mandate/access-control issue in this epic) to record the assigned class.
- No agent can be marked "production" in the agent registry without a recorded classification.
- A named approver signs off on the classification before first deployment; the approval is captured in the registry entry.
- Class 4 (prohibited) results in an explicit rejection, not a deployment path.

## Dependencies

Builds on #75 (agent-registry / `.agent-card.yml`). Feeds the approval-gate and mandate work in this epic.

## Prerequisites

⚠️ Prerequisite: Agreement from BOD leadership on who holds sign-off authority for classification approvals — Owner: Project Director.