[EPIC CAG] Enterprise coding-agent governance MVP calibration

## Why this matters

A conversation with a Tech Director and an FDE surfaced a concrete enterprise need: teams want organization-level rules and execution boundaries for coding agents, but some cannot adopt one vendor's enterprise plan and others deliberately use Codex, Claude Code, Cursor, or more than one of them.

Those teams otherwise have to assemble instruction distribution, sandboxing, permissions, identity, approvals, and audit separately for each vendor. The result is expensive, inconsistent, and easy to overstate.

This Epic calibrates the Agenova MVP around that real scenario without turning Agenova into a coding-agent-only product.

## Target scenario

One enterprise can run a governed coding assignment through more than one coding-agent vendor while preserving the same Agenova meaning:

```text
trusted principal + company rules + ClaimRequest
-> effective authority and one SandboxClaim
-> isolated filesystem and governed external access
-> vendor-specific agent integration
-> consistent decisions and evidence
```

Company Instructions may be materialized in vendor-native forms, while hard Policy remains deterministic and external to the agent. A user may add more specific task or repository guidance but cannot expand the organization's hard authority ceiling.

## Existing Agenova foundation

This Epic extends rather than duplicates the existing MVP:

- E1 defines backend-neutral AgentTemplate, ClaimRequest, SandboxClaim, decision, and evidence contracts.
- E2 resolves hard Policy and EffectiveAuthority.
- E3/E8 own runtime lifecycle and real-backend isolation evidence.
- E4 owns Tool and Model Gateway enforcement.
- E5 owns the single evidence model.
- E10-T5 (#74) provides the product-discovery evidence behind this calibration.

## Child outcomes

- managed Instructions delivery and the Skill boundary;
- the minimum agent filesystem boundary;
- one bounded ApprovalRequired interruption/resume path;
- upstream identity and trusted metadata integration;
- platform-agnostic coding-agent adapters and a multi-vendor POC.

## Definition of Done

- Each child records what enters the MVP, what is reference-only, and what is explicitly deferred.
- At least one reproducible coding-agent scenario proves Agenova's existing claim, authority, backend, gateway, and evidence semantics rather than inventing a parallel coding governance model.
- The selected POC demonstrates how the same company intent maps to at least two vendor surfaces, or records a concrete blocker with evidence.
- Coding-specific behavior remains inside example artifacts or agent-integration adapters; shared contracts remain useful for non-coding agents.
- Any accepted MVP scope change is reflected in the PRD and architecture contract through a reviewed implementation ticket.

## Scope guard

This Epic does not commit a full enterprise administration suite, Web IDE, managed workspace service, IdP/MDM product, workflow engine, or production certification against hostile-agent bypass.

Start by reading the existing E1, E2, E3/E8, E4, E5, and E10-T5 (#74) foundations, then review the PRD and architecture contract. Break the child outcomes into reviewed implementation tickets, recording MVP, reference-only, and deferred scope; done requires a reproducible multi-vendor scenario or an evidenced blocker without changing shared contracts.

Written by the indexing model from the issue text.