Summary

The task previously hydrated full ObservationSpan instances — including
large attribute and I/O payloads — before dispatching evaluations. On
eval tasks covering a high volume of spans this caused worker OOMs.

This change limits the query to span IDs and removes redundant
materialization:

• Restrict fetched columns via .only("id") and values_list("id", flat=True)
  in place of full model instances.
• In the sampling path, reuse the IDs returned by the random-sample query
  directly rather than issuing a second filter(id__in=...) round-trip.
• Apply the cnt cap by slicing the queryset (list(qs[:cnt])) so the LIMIT
  is pushed to Postgres instead of forcing full evaluation via len().
• Dispatch celery tasks by iterating span IDs directly.

Behavior is unchanged: the same set of span IDs is enqueued for
evaluation and persisted to spanids_processed.

Type of change

• 🐛 Bug fix (non-breaking change which fixes an issue)
• ✨ New feature (non-breaking change which adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📖 Documentation only
• 🧹 Chore / refactor (no user-visible change)
• 🚀 Performance improvement
• 🧪 Test-only change

How was this tested?

Screenshots / recordings (if UI)

Checklist

• My code follows the style guide
• I've added tests that prove my fix is effective or that my feature works
• make check-all / yarn check-all passes locally
• I've updated the documentation where relevant
• No hardcoded secrets, URLs, or PII
• I've signed the CLA