c744adadfd
Third run with the inducer on from flight start went 900s without freezing, so the tally is 2 for / 1 against. Stated confound: the inducer was far weaker this run (~47-50s per CPU pass vs ~4.2s in run 2, because two subagents were saturating the box -- 22 passes in 15 min instead of ~150). That is consistent with either reading, so it is not scored as a confirmation. What it does establish is that the inducer is not sufficient, and that future runs must report its measured rate rather than just that it was on. Second healthy stability sample: 9 stable / 14 vary vs 12 / 13 in the first, intersecting to only 8 threads. So the stable set is run-dependent and must be measured within the run it is used in -- the same lesson as gdb thread numbering, one level up. The frozen half of the distribution experiment is still uncollected.