5addf6b7

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

document GL638 precision escalation study

Record 100 actual optimized-LMB escalations, ordinary and forced-Arb
replays, selected cut diagnostics, and conditional small-weight waiver
counts. Document the unavailable estimator gate and signed/absolute
reference mismatch without changing numerical source.

User request:
Do an experiment completely separate from the current two runs currently active (do not disrupt them) where you run the optimized\_lmbs setup integration (on the different state and integration workspace folder) and record 100 points that required precision escalation (use revertible temporary modifications  in gammaloop for this if need be) and study their nature (by e.g. replaying them with inspect) and determining the source of the instability (if you can) and also whether any reasonable threshold of `min_abs_wgt_for_escalation` could have helped.
Also verify that the current active runs are not writing massive amount of info into log files (it shouldn't, but good to check). If it does, adjust the setup to prevent that, and kill the runs and refresh+restart the integrations.
Once the above is completed, deem the goal reached and summarize your findings.

Follow-up:
And did you see the new option `min_abs_wgt_for_escalation` set at 1e-12 helping prevent as much escalation for the lmb\_optimized run?

Timeline