feat(plugin): trial output-consistency metric via embedding similarity

[christso]

Canonical Plan

This issue body is the source of truth for implementation.

Objective

We introduce a trial-output consistency metric as a plugin/reference capability.

Why This Matters

• Trial score variability does not directly describe semantic output consistency.
• We need a dedicated signal for response stability across repeated attempts.
• This unlocks stronger diagnostics for non-deterministic agent behavior.

Why This Location

• Consistency scoring method choice is specialized and likely to evolve.
• Plugin-first lets us experiment without hardcoding a narrow built-in.
• This aligns with AgentV’s extensibility-first architecture.

Architecture Boundary

Plugin-first (aligned with AgentV principles).

Deliverable Location

• Primary location: examples/features/trial-output-consistency/
• Plugin/judge location: examples/features/trial-output-consistency/judges/
• Runnable eval location: examples/features/trial-output-consistency/evals/
• Usage/docs: examples/features/trial-output-consistency/README.md

Design Latitude

We can choose embedding backend, similarity strategy, and artifact flow, provided the solution stays extension-oriented.

Acceptance Signals

• We provide a runnable reference implementation that computes consistency across repeated trial outputs.
• We expose the result as a named metric in evaluation workflows.
• We define explicit behavior for low-trial and edge-case inputs.

Non-Goals

• No built-in core evaluator unless plugin-first proves insufficient.