feat(phoenix): run real AgentV targets inside Phoenix experiments

[christso]

Objective

Replace synthetic Phoenix adapter task outputs with actual AgentV target execution so Phoenix experiments represent real AgentV behavior.

Acceptance Signals

• Phoenix experiment tasks invoke AgentV execution for the corresponding normalized test case.
• Phoenix task output matches the actual AgentV target output, not synthesized expected-output content.
• AgentV scores, assertions, verdicts, duration, cost, token usage, target, and stable agentv_test_id metadata are preserved in Phoenix run/evaluation metadata.
• Missing target/configuration produces a clear run error and is not reported as an evaluator failure.
• Dry-run/reference mode remains network-free and clearly separated from live execution.
• At least one live Phoenix smoke against a deterministic example records real task runs and evaluator runs.

Implementation Notes

Keep AgentV eval YAML, target execution, workspace lifecycle, and scoring authoritative. Phoenix should host experiment artifacts, not become a parallel AgentV runtime.

Relevant files:

• packages/phoenix-adapter/src/phoenix/run-experiment.ts
• packages/phoenix-adapter/src/run/options.ts
• packages/phoenix-adapter/src/run/run-suite.ts
• packages/phoenix-adapter/src/agentv/load-spec.ts
• packages/phoenix-adapter/src/phoenix/types.ts
• packages/phoenix-adapter/test/phoenix-run-experiment.test.ts
• packages/phoenix-adapter/test/agentv-execution.test.ts

Non-goals

• Do not reimplement workspace pooling, Docker lifecycle, matrices, or trials inside the adapter.
• Do not make Phoenix required for normal agentv eval execution.