feat(tracing): stamp sgp evals run and row ids on SGP spans - #545

OVE-1253. Pairs with an sgp-evaluations PR that tags each generation-unit task.

The problem. Spans of an evals generation-unit task carry no run or row id, so a spans search cannot find them.

The fix.

• Stamp: the SGP copy of each span of an eval task gets flat run, row and attempt keys. A task that loses the eval marker stops being stamped.
• Sync and async agents: the ACP server reads the task it receives.
• Temporal agents: the ids travel through the workflow memo to the worker's activities, child workflows included. A span keeps the ids its task had when it started, so a reused task id cannot strip a queued span. Not covered: spans whose trace_id and task_id are both not the task id.

Test plan

• Real Temporal dev server run: eval workflow stamps the three keys, a plain one stamps nothing
• Real Temporal dev server run with a child workflow: its activity is stamped with the parent task's ids (unstamped before this change), and a history recorded with the old interceptor replays with no nondeterminism

Changes since review

• A span's run and row ids are fixed when it starts, so a task id reused before export cannot change them.
• A span that started with no eval task can no longer pick up a later run's ids.
• A failed upload keeps the ids, so a retry sends the same span.
• Activities of a child workflow carry the ids of the task that started it.

Confidence Score: 4/5

This PR is not ready to merge because an early plain span can be saved with a later eval run’s IDs.

Findings

1.  Plain spans get eval labels ▶

### Issue 1
src/agentex/lib/core/tracing/sgp_evals.py:95-96
If a plain span starts before an SGP processor exists, `capture_for_span` saves no empty marker. If an eval task later takes the same task ID, that earlier span can export with the eval task’s run and row IDs. It then appears in searches for work it did not do. Keep the empty result from the span’s start.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Eval-marked tasks now add flat run, row, and attempt ids to their SGP span copies, making spans searchable by evaluation. The PR wires that attribution through ACP handlers and Temporal workers, while keeping queued spans tied to the task metadata they had when they started.

• SGP span copies show which eval run, row, and attempt they belong to.
• Temporal workers carry eval ids from workflows into activity spans.

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Task[Eval task metadata] --> Memo[Workflow memo]
  Memo --> Header[Activity headers]
  Header --> Registry[Worker task IDs]
  Task --> Registry
  Registry --> Capture[Span-start capture]
  Capture --> Export[SGP span export]
  Registry -->|No capture saved| Export

_{Reviews (6) · Last reviewed commit: "style: sort interceptor imports" · Reviewed by Greptile}