Improve OTLP telemetry span names with dataset context #58

Objective

Improve the readability of OTLP evaluation spans by including dataset row context (input text) in span names and attributes.

Current Behavior

Item spans are named generically:

• eval_item 1, eval_item 2, eval_item 3

This makes it hard to identify which dataset row a span corresponds to when viewing traces in AI Toolkit, Jaeger, or other collectors.

Desired Behavior

Item spans should include a truncated version of the input query:

• eval_item 1: What is the weather in Seattle today?
• eval_item 2: Convert 100 USD to EUR

And the input_text and expected_text attributes should be populated on item spans.

Implementation Notes

• Thread dataset rows (or at minimum input_field values) from the runner to _emit_item_spans()
• Pass input_text to eval_item_span() — the parameter already exists but is not used by the caller
• Truncate input text in span name to ~80 chars to avoid excessively long span names
• The eval_item_span() function in utils/telemetry.py already accepts input_text and expected_text — only the caller in services/runner.py needs updating

Files to Change

• src/agentops/services/runner.py — pass dataset row data to _emit_item_spans
• tests/unit/test_telemetry.py — update tests if span name format changes

Related

• PR Add OTLP tracing support and documentation for evaluation runs #56 (OTLP tracing feature)
• PR feat: extend Foundry cloud evaluator coverage to 22 built-in evaluators (#51) #57 (evaluator expansion — issue Review CI/CD Based on Field Insights #51)