[Bug]: promptDefenseGrade does not reflect operator configuration

## Package

Other — `agent-governance-opencode`

## Description

`getPolicyStatus` reports `promptDefenseGrade: "A"`, `coverage: 12/12`, and no missing defenses across materially different operator configurations. `minimumPromptDefenseGrade` sits alongside this status in the policy schema, which makes the grade look like assurance about the supplied configuration even though it appears to describe a constant built-in detector catalogue.

Observed configurations all returned the same result:

| configuration | grade | coverage |
|---|---:|---:|
| no `additionalContext` key | A | 12/12 |
| `additionalContext: []` | A | 12/12 |
| `additionalContext: ["probe"]` | A | 12/12 |
| explicitly unsafe context instructions | A | 12/12 |

`promptDefenseBlocking` remained false throughout.

## How does this impact your work?

The session-start line prints `promptDefense=A`, and users may cite it as evidence that their configuration is hardened. We currently label it “advisory only; independent of configuration” and do not use it as a gate.

## Steps to Reproduce

1. Load policies with the four `additionalContext` variants above.
2. Call `getPolicyStatus` for each policy.
3. Compare `promptDefenseGrade`, coverage, missing defenses, and `promptDefenseBlocking`.

## Suggested resolution / acceptance criteria

Either:

- calculate a grade that reflects the operator-supplied configuration and enforce `minimumPromptDefenseGrade`; or
- rename the status field and document that it reports built-in catalogue coverage rather than configuration assurance.

The session-start log and status response should make the distinction unambiguous, with regression tests for materially weak and strong configurations.

## Environment

- `@microsoft/agent-governance-opencode@5.0.0`
- Node 26.7.0
- macOS arm64
- Verified against AGT commit `81955d4`

## Code of Conduct

- [x] I agree to follow the Microsoft Open Source Code of Conduct

## Coordination status

- Active implementation: #3677.
- Related follow-up: #3669 covers session-scoped state and staged evaluation; it is not implemented by #3677.

Please coordinate on #3677 rather than open a parallel fix for this issue.

Start with getPolicyStatus and the session-start log described in the issue, then review active implementation #3677 rather than opening a parallel fix. Done means the status and log clearly distinguish configuration assurance from built-in catalogue coverage, with regression tests for materially weak and strong configurations.

Written by the indexing model from the issue text.