You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fix(scripts): derive harness validator bounds instead of hard-coding them
Smoke-running the pipeline end to end surfaced two literals that were only
correct for the full profile. The calibration check demanded 590 ps samples and
the workload validator demanded exactly 30 turns, neither derived from the
constants that actually determine them. A literal like that is not merely
brittle: change the observation window or the wave count and the validator
starts rejecting correct runs, or worse, accepting truncated ones.
Sample floor now comes from OBSERVE with two percent slack for scheduler jitter,
and turn counts come from CLIENTS times WAVES. Real-run values are unchanged.
Adds a smoke mode behind OCX_RSS_HARNESS_SMOKE that shortens durations only and
stamps every artifact smoke:true valid:false, so a shortened run can never be
read as evidence or compared against a real one. It exists because a seven-hour
session is a poor way to discover that the plumbing does not connect -- and it
immediately paid for itself by finding both literals above.
0 commit comments