feat(phoenix): complete AgentV Phoenix integration

[christso]

Objective

Track the remaining work to complete the AgentV Phoenix integration after the initial adapter and Phase A OTel preset work.

Context

Draft PR #1280 is the seed/handoff PR. It adds the planning document and first supported user-facing slice: agentv eval --export-otel --otel-backend phoenix plus documentation. It does not close this parent issue.

This parent issue covers the follow-up implementation issues needed before the Phoenix integration can be called complete.

Scope

Complete the integration as two intentionally bounded surfaces:

• Phoenix OTLP observability for normal AgentV eval runs.
• A Phoenix dataset/experiment adapter that keeps AgentV eval YAML and AgentV scoring semantics authoritative.

Subissues

Recommended sequencing:

1. feat(phoenix): complete deterministic adapter parity #1282 - Deterministic adapter parity.
2. fix(phoenix): make full dry-run structural parity actionable #1283 - Full dry-run structural parity.
3. feat(phoenix): run real AgentV targets inside Phoenix experiments #1284 - Real AgentV target execution inside Phoenix experiments.
4. feat(phoenix): support AgentV-authoritative LLM graders and rubrics #1285 - AgentV-authoritative LLM graders and rubrics.
5. feat(phoenix): support trace and metric graders through Phoenix trace IDs #1286 - Trace and metric graders through Phoenix trace IDs.
6. chore(phoenix): decide and implement adapter package publishing posture #1287 - Adapter package publishing posture.

Acceptance Signals

• Users can use the documented Phoenix OTel preset for normal eval observability.
• Adapter support gaps are explicit and conservative.
• Phoenix experiments run real AgentV targets before claiming target parity.
• AgentV scoring semantics remain authoritative for AgentV-authored evals.
• Full dry-run parity is green or has documented exclusions.
• Live Phoenix verification covers both OTLP export and at least one experiment path.
• Maintainers can tell whether the adapter is intentionally private or publishable.

Non-goals

• Do not replace AgentV's local result JSONL/artifact model with Phoenix.
• Do not make Phoenix a required dependency for normal eval execution.
• Do not reimplement workspace lifecycle, Docker sandboxing, target matrices, trials, or custom assertion discovery inside Phoenix unless a later concrete issue proves it is required.