Improve OpenAI evaluator telemetry coverage with gpt-5.5 medium

[mltseva]

id: 589e6fac-7d92-4f60-aad6-6c9d559e7a67
artifact_path: experiments/tracy-openai/gpt-5.5-medium-2/
providers: ["openai"]
status: succeeded
score: 87
checks: 1522 / 1794
passed scenarios: 15 / 86
provider_errors: 4

Category summary:

openai/audio avg 96.9
openai/batches avg 95.0
openai/conversations avg 94.2
openai/files avg 93.5
openai/moderations avg 92.3
openai/videos avg 96.0
openai/embeddings avg 86.3
openai/responses avg 83.4
openai/images avg 78.0
openai/chat avg 66.7

Worst scenarios:

26 openai/chat/lifecycle provider_error=true
26 openai/chat/vision provider_error=true
40 openai/responses/streaming
45 openai/chat/streaming
60 openai/chat/tools
63 openai/responses/invalid_empty_input
68 openai/images/streaming
69 openai/images/variation provider_error=true

Main remaining gaps:

• Operation/API type regressions are the biggest issue in this artifact:
  • gen_ai.operation.name expected='generate_content' actual='responses' 25 times.
  • openai.api.type expected='chat_completions' actual='chat' 12 times.
  • gen_ai.operation.name expected='chat' actual='chat.completions' 10 times.
• Numeric request values are still emitted as strings in some places, e.g. tracy.request.n actual='1'.
• Streaming and lifecycle scenarios still miss some usage/status/id fields.
• Provider errors remain in:
  • openai/chat/lifecycle
  • openai/chat/vision
  • openai/images/variation
  • openai/videos/lifecycle