A cap checked once per tick, and a tick that took three seconds
Both orbit runs went three seconds past their 110 second limit. The limit was checked between ticks, and the last tick scanned 64 MB and then slept.
The entry point gives each policy 110 seconds. On the FlatSat the shrinking quorum ran for 113.2 and fixed TMR for 110.0. In orbit, 113.5 and 113.1.
The cap was checked once per tick, at the top. A tick that started a scan of the 64 MB sensor block just under the cap ran the whole scan, which is slow on a Cortex-A9, and then took its throttle nap. At a 5% CPU share the nap is nineteen times the work. Fixed TMR’s last tick landed on a scan in orbit and not on the FlatSat, which is why one host showed 110.0 flat and the other did not.
The scan now takes a deadline and checks it every 1 MiB, a few milliseconds
of work. Past the deadline it returns false, the run ends with
stopped,time-limit on that tick, and the partial scan is not counted. The
nap is clipped to the time left. Scenario 26 scans 32 MB on every tick with a
one second cap and requires the wall time under 1.5 seconds. It ends at 1.001.
The fix is in v0.4.0, which has not flown. The binary that flew has the three seconds in it, and the research entry calls that the one defect.
The prediction said about 225 seconds and the run took 227. The overrun was inside the prediction, because the FlatSat had it too.