zli train runs far past --max-time-secs. The overrun happens in the "Training ACE graph N/M" phase, which appears not to check the deadline; earlier phases respect it.
Repro (built from a9de25e, macOS arm64, 8 cores):
# sample dir: 3 CSV files, ~8 MB each (any similar samples reproduce it)
zli train --profile csv samples/ -o out.zlc --max-time-secs 120 -f
Observed, all with a 120 s budget, all exit 0:
--profile csv, dataset A (NYC 311 export): 495 s and 504 s wall (two runs)--profile csv, dataset B (Chicago crime export): 543 s wall--profile serial: 146 s wall- With
--no-ace-successors, the budget holds (60 s budget → 49 s wall)
The progress bar makes it visible: the deadline passes while it cycles "Training ACE graph N/15".
Expected: either the ACE phase checks the deadline (stop early, return current best), or the help text states the budget is advisory and excludes the ACE phase.
Can share the exact sample files if useful.
zli trainruns far past--max-time-secs. The overrun happens in the "Training ACE graph N/M" phase, which appears not to check the deadline; earlier phases respect it.Repro (built from
a9de25e, macOS arm64, 8 cores):Observed, all with a 120 s budget, all exit 0:
--profile csv, dataset A (NYC 311 export): 495 s and 504 s wall (two runs)--profile csv, dataset B (Chicago crime export): 543 s wall--profile serial: 146 s wall--no-ace-successors, the budget holds (60 s budget → 49 s wall)The progress bar makes it visible: the deadline passes while it cycles "Training ACE graph N/15".
Expected: either the ACE phase checks the deadline (stop early, return current best), or the help text states the budget is advisory and excludes the ACE phase.
Can share the exact sample files if useful.