3f942912

By: ValentinHirschi <valentin.hirschi@gmail.com>

handle fully weighted underflow in stability checks

Recognize a suppressed averaged component using its mean absolute fully
weighted probes below the binary64 normal boundary. Preserve native values
and relative diagnostics, retain component isolation and amplification guards,
and cover the criterion with focused tests. Record the twelve 600 GeV GL638
screens and precise maximum attribution; c=4 is an exploratory candidate.

Validation: cargo check and clippy passed; six focused tests passed. The exact
4096-draw O2 regression has zero unstable/NaN samples and unchanged estimates,
errors, maxima, channel data and precision distribution. Larger fresh-seed
calibration is ongoing and has separate stability flags under investigation.

User instructions:
Yes, update this requirement, f64 underflow can happen and can be sent to zero. Of course this may mean that stability test trivially pass, but it is in this case anyway very small because such small magnitude can only come from exp functions far down the tale. So it's good, you can udpate this and also by yourself take that decision in the future if it happens.

Continue as planned, but make sure that before the long run, you do multiple shorter 15 min ones with 50 cores and different channels choices and integration setup, each time investigating the resulting max weight.
Go for the longer hours-long run when you know you have a setup that is likely to give a <10% error run on the real part. Start with sqrtS of 600 GeV. It may be that the central value is accidentally close to zero, but the integral of the absolute value should really converge well then.

Timeline