feat(phoenix): support trace and metric graders through Phoenix trace IDs

[christso]

Objective

Enable trace-derived AgentV graders in Phoenix experiments once Phoenix spans can be associated with each AgentV test case.

Acceptance Signals

• Phoenix task/evaluator context exposes a stable trace ID or span lookup path for each test case.
• Phoenix spans can be translated into AgentV trace summary or metric inputs.
• At least one tool-trajectory happy path and one execution-metrics threshold are supported.
• Missing trace IDs produce clear unsupported/failed explanations rather than empty passes.
• Phoenix API lookup failures are surfaced as evaluation errors with clear guidance.
• Live verification demonstrates trace-derived scores from Phoenix-ingested span data.

Implementation Notes

Build this after real AgentV execution is wired into Phoenix experiments. Start narrow with one tool trajectory and one metrics case before broadening.

Relevant files:

• packages/phoenix-adapter/src/phoenix/run-experiment.ts
• packages/phoenix-adapter/src/phoenix/types.ts
• packages/phoenix-adapter/src/evaluators/registry.ts
• packages/phoenix-adapter/test/evaluators/trace-metrics.test.ts
• packages/core/src/evaluation/trace.ts
• apps/cli/src/commands/inspect/utils.ts

Non-goals

• Do not support trace graders without a reliable trace ID/span association.
• Do not silently pass when metrics are missing.