Track NNsight tracing interference with coverage.py

[speediedan]

Summary

We have a documented coverage interference issue when nnsight tracing is exercised under coverage.py.

The current note in distributed-insight/project_admin/interpretune/adapter_reference/NNsight/tracing_coverage_issue.md points to nnsight calling sys.settrace(None) inside its tracing implementation, which likely wipes out coverage's trace function during with model.trace(...) execution.

Current evidence

• The issue appears to be tied to tests that execute actual NNsight tracing contexts.
• The local note identifies these known trigger cases:
  • TestNNsightTracing::test_nnsight_trace_context
  • TestNNsightRemoteExecution::test_remote_local_consistency
• The likely upstream touchpoint is nnsight/intervention/tracing/base.py where trace handling resets the interpreter trace function.

Why track this

• This can undercount coverage for adapter and analysis code paths that are actually exercised.
• It makes coverage interpretation less reliable exactly where we are adding NNsight-backed functionality.
• We should decide whether to work around it locally, isolate affected coverage expectations, or push for an upstream fix.

Follow-up

• Reproduce the coverage drop with a minimized pytest --cov case.
• Confirm the exact nnsight version and upstream line(s) involved.
• Open or link an upstream nnsight issue once the repro is reduced.
• Decide whether affected tests should remain in standalone/optional buckets until the coverage interaction is understood.