[gipert]

Summary

• Explicitly del tb_data, dsp_data before the second lh5.read call in par_geds_dsp_nopt so the first read's memory is released before the second allocation, reducing peak RSS by ~2×
• Fix no-plot branch passing raw_files instead of tb_data to noise_optimization
• Extract n_events = opt_dict.pop("n_events") before the del so the value is available for the second read

Test plan

• Run par-geds-dsp-nopt on a test channel and verify it completes without memory errors
• Verify --plot-path and no-plot-path branches both produce correct output

[gipert]

@ggmarshall what do you think about this? please close if wrong

[ggmarshall]

don't think this does much, I think I need to update the pygama function

[ggmarshall]

But I would include anyway