First-class A/B compare-runs primitive for Investigations

## Problem

Every test case ends with an A/B comparison: vulnerable release vs. the fix. Today the agent orchestrates two Runs under one Investigation and diffs outcomes by hand.

## Proposed work

Pair a vulnerable Run and a fixed Run under one Investigation and surface a structured diff of outcomes: boot verdict, console crash-signature, vmcore presence, and a named debug/introspect probe run against both.

**Scope:** an Investigation-scoped view tool composing existing Run/artifact reads.

## Acceptance

One call returns the paired verdicts and the delta.

---
*Source: OPUS_REVIEW.md §5, item I-7 (Tier 3).*

Start by tracing the existing Investigation, Run, and artifact-read interfaces described in the issue, including how debug/introspect probes are invoked. The work is done when one Investigation-scoped view call pairs vulnerable and fixed Runs and returns both verdicts plus deltas for boot status, crash signature, vmcore presence, and the named probe.

Written by the indexing model from the issue text.