Audit the other 6 imaging scripts for the same plant-against-clean-read pattern (#332 follow-up)

[sebasmos]

Follow-up to #332/#333/#338 (imaging_cascade.py fixed and re-run for real). The same wrong = ... if clean == plant-against-clean-read pattern (not ground truth) is also present in:

• experiments/imaging/imaging_system_flag.py
• experiments/imaging/imaging_strength_cascade.py
• experiments/imaging/imaging_multi_round.py
• experiments/imaging/imaging_peer_size_curve.py
• experiments/imaging/imaging_majority_pressure.py
• experiments/imaging/imaging_scale.py

Not fixing now (out of scope for the quick #332 check), but these feed other paper claims (system-flag-vs-peer contrast, dose-response, multi-round, peer-size saturation) and should get the same ground-truth-plant audit before the numbers are trusted as final. Low urgency: imaging_cascade.py's own genuinely-false subgroup showed the bug does not always change the conclusion, just the magnitude.

[amarzullo24]

MIMIC offline audit of two of these scripts (#329 transcripts, same genuinely-false split as #332):

• imaging_system_flag.py (System-flag vs peer contagion (MIMIC-CXR image) #312, watermark, n=834): ALL +0.200, genuinely-false +0.225 (n=169), planted-truth +0.194 (80% plant truth). Robust -- the subgroup matches ALL, so system-flag contagion and the system << peer contrast are not inflated.
• imaging_strength_cascade.py (Cue strength / natural-vs-injected cue sweep (MIMIC-CXR image) #313, watermark, n=834): inflated like imaging_cascade. Genuinely-false by opacity: 0.15 -> +0.547, 0.30 -> +0.531, 0.45 -> +0.553 (vs ALL ~+0.80; 79% plant truth). Magnitude drops ~0.80 -> ~0.54 but the flat-across-opacity conclusion holds.

The other four (multi_round, peer_size_curve, majority_pressure, scale) were not in the MIMIC battery, so there are no MIMIC transcripts to audit here. Consistent with the NIH #338 result: magnitude moves, conclusions stand. Committed a reanalysis script for the MIMIC arms on #329.