Minor reporting bug, no scoring impact as far as I can tell, but it’s confusing when you’re tuning against the λ residual penalty.
Under --runner subprocess, the aggregate budget-breakdown panel prints:
Residual Wall Time [residual_wall_time_s] 39806.906344s
(~11 hours) for a run whose wall clock was a couple of minutes, while the per-estimator panel in the same output reports a sane ~0.05 s. Looks like the aggregate is summing across
subprocess-lifetime contexts, or a timebase/units mix-up, rather than reporting estimator residual.
Scoring appears to use the per-estimator number (our multiplier and score came out as expected), so this seems cosmetic — but the 11-hour figure in the panel misleads anyone reading it
to budget their residual.
Notably the value is bit-for-bit identical across independent runs (39806.906344s every time), which suggests it’s an uninitialized/sentinel field being displayed rather than an
actual accumulated wall time.
Repro: any estimator, mini split, --runner subprocess; compare the two “Residual Wall Time” lines in the output.