[aclerc]

Summary

The three filter-stage operating-curve figures (ops_curves.png, ops_curves_kept_only.png, ops_curves_excluded.png) were drawn for the test turbine only. They are now also written for every reference turbine in the run, from that reference's own power, wind speed, pitch and generator rpm, as {ref}_ops_curves.png, {ref}_ops_curves_kept_only.png and {ref}_ops_curves_excluded.png in the same plots/2_filter/ folder (the per-turbine prefix convention {wtg}_coverage_timeseries.png / {wtg}_pairing_gap.png already use). The test turbine's filenames are unchanged. Each figure's suptitle names the turbine it shows.

Why

In a toggle run a reference with bad data (a stuck anemometer, a mis-scaled power signal, a special operating mode nobody flagged) corrupts the estimate as surely as a bad test turbine, but the test turbine's figures cannot show it. With one set of figures per reference the whole reference group can be eyeballed in the filter stage before the estimate is trusted.

• Kept is the method's used timestamps for every turbine (a timestamp is used only when the test turbine and every reference passed), so a reference's kept-only figure shows exactly which of its rows fed the estimate.
• The excluded figure colours each reference by its own ColumnSchema.exclude_row flag (new DiagnosticContext.turbine_excluded_mask), not the run-wide excluded_ts mask, which is the union over turbines and would hide which turbine carried the flag. A reference with nothing excluded gets no excluded figure.
• A reference with too little wind speed (or no power column) gets no figures, as the test-turbine code already did.
• Plots in common._PLOTS may now return Path | list[Path] | None; write_common_diagnostics flattens the list form.

Test plan

• New TestReferenceOpsCurves in tests/benchmarking/diagnostics/test_diagnostics.py (written first, watched fail, then made green): the three files are written per reference into 2_filter; the test turbine's filenames are unchanged and never prefixed; a reference's excluded figure follows its own flag (a flagged reference gets one, an unflagged one does not, while the run-wide mask is set); turbine_excluded_mask reads the per-turbine column; no reference excluded figures without an exclusion concept; a reference with no wind speed draws nothing for itself.
• uv run python -m pytest tests/benchmarking/diagnostics -q : 30 passed
• uv run python -m pytest tests/benchmarking/baselines/test_toggle_specialist.py -k "Exclu or exclu or plot or diagnostic" : 28 passed
• uv run poe lint (ruff format, ruff check, mypy) : clean

🤖 Generated with Claude Code

[Copilot]

🔵 Needs a closer look

Diagnostics should use the method’s effective reference set rather than every non-test turbine in the input data.

No open issues 🎉

What changed in this PR

This pull request extends filter-stage operating-curve diagnostics from the test turbine to reference turbines, using each turbine’s own signals and exclusion flags while preserving existing test-turbine filenames.

Changes:

• Generates standard, kept-only, and excluded operating-curve figures for reference turbines.
• Adds per-turbine exclusion-mask handling.
• Supports diagnostic plots returning multiple output paths.
• Adds regression coverage for reference figures, filenames, exclusions, and missing signals.
• The effective reference set used by the estimate is not currently propagated into diagnostics, so figures may be generated for turbines outside the actual reference group.

File	Description
tests/​benchmarking/​diagnostics/​test_diagnostics.py	Tests reference figures, filenames, exclusion behavior, and missing-signal handling.
benchmarking/​diagnostics/​curves.py	Generalizes operating-curve plotting across the test turbine and references.
benchmarking/​diagnostics/​context.py	Adds per-turbine exclusion-mask extraction aligned to the diagnostic index.
benchmarking/​diagnostics/​common.py	Supports diagnostic functions returning lists of output paths.

_{Review effort: Lite}