feat(tooling): split combined results JSONL by target (external-first)

[christso]

Canonical Plan

This issue body is the source of truth for implementation.

Objective

We improve multi-model benchmarking ergonomics by producing per-target result artifacts from combined JSONL outputs.

Why This Matters

• Multi-target runs are harder to compare quickly when all records stay in one combined file.
• Teams repeatedly write ad-hoc filtering scripts before they can run model-vs-model analysis.
• Standardizing this utility reduces operational friction in benchmark workflows.

Why This Location

• This is an output-shaping concern, not scoring/execution behavior.
• We can solve it as tooling without changing core evaluator/runtime semantics.
• Keeping it external-first preserves AgentV’s lightweight-core design.

Architecture Boundary

External-first. We prefer wrapper/tooling and avoid core evaluator/runtime changes.

Deliverable Location

• Primary location (in-repo tooling path): examples/features/benchmark-tooling/
• Script location: examples/features/benchmark-tooling/scripts/ (for split-by-target utility)
• Usage/docs: examples/features/benchmark-tooling/README.md

Design Latitude

We can choose the exact utility interface and filename strategy as long as output is deterministic and easy to use.

Acceptance Signals

• We can derive one deterministic JSONL per target from a combined results file.
• We handle target names that require safe filename normalization.
• We document the downstream compare workflow.

Non-Goals

• No runtime scoring semantics changes.
• No mandatory schema changes for existing result records.