fix(phoenix): make full dry-run structural parity actionable

[christso]

Objective

Resolve or explicitly exclude current Phoenix adapter full dry-run failures so the dry-run report can become a reliable regression signal.

Acceptance Signals

• Full dry-run exits successfully, or reports only explicitly documented non-blocking exclusions.
• The four known failures are investigated and either fixed at source/baseline level or excluded with clear reasons:
  • examples/features/matrix-evaluation/evals/dataset.eval.yaml
  • examples/features/prompt-template-sdk/evals/dataset.eval.yaml
  • examples/features/tool-trajectory-simple/evals/dataset.eval.yaml
  • examples/features/weighted-graders/evals/dataset.eval.yaml
• Parity report output distinguishes conversion crashes from source/baseline drift and documented exclusions.
• packages/phoenix-adapter/docs/e2e-verification.md reflects the current expected full dry-run result.

Implementation Notes

Prefer fixing stale baselines or source references over adding new mechanisms. If an eval intentionally diverges, encode a small documented exclusion in adapter parity reporting.

Relevant files:

• packages/phoenix-adapter/src/parity/compare.ts
• packages/phoenix-adapter/src/parity/report.ts
• packages/phoenix-adapter/test/parity.test.ts
• packages/phoenix-adapter/docs/e2e-verification.md
• The affected example evals/baselines listed above

Non-goals

• Do not accept both old and new wire key names unless the old form has shipped and needs compatibility.
• Do not turn dry-run into a live Phoenix or live provider test.