docs: Explain how reviewer tool evidence affects resumed reviews

[zzwong]

Small documentation follow-up to #540, where the runtime fix landed and the artifact documentation was explicitly deferred.

The useful thing to explain is why a reviewer task can succeed while its review still cannot approve. For example, this fragment of metadata.json is possible:

{
  "status": "succeeded",
  "reviewer_tool_evidence": {
    "diff_status": "failed",
    "diff_diagnostic": "fixed diff unavailable"
  }
}

Here, succeeded means the task produced valid structured output. The failed diff tool still makes reviewer coverage incomplete_tool and clamps an otherwise approving result to comment; the existing regression test covers this. The artifact guide currently omits reviewer_tool_evidence, so this distinction requires reading the code.

A focused update to docs/llm-task-artifacts.md should be enough:

• Describe diff_status (not_invoked, incomplete, succeeded, failed) and optional diff_diagnostic, following the existing type definitions.
• Explain its effect on coverage and approval, and why cached tasks restore it from metadata. That gives the guide a concrete explanation of why the field is behaviorally significant.
• Describe the current version-1 compatibility behavior: an absent field means no adapter-provided evidence; it is distinct from an explicit not_invoked result. Without evidence, the normal coverage checks still apply.

Done when the guide explains the example above, preservation on resume, and the missing-field case. A short schema entry and example can close this documentation follow-up.