Repository navigation
v2.13.3
What this cut is
A selftest control that went red on the machine's state rather than the tree's says what it could not read instead. Every file build writes is byte-identical to v2.13.2's, and no input format, error message or assertion moved; nothing a consumer runs changed.
TY40's planted unit under memory pressure is a SKIP that states both figures, never a FAIL (#1185 → #1197). The control starts a unit holding a stated amount of memory and expects the run to name it by its figure; on macOS the figure read had twice been lower than what the unit held, and the control failed on an unchanged tree. Measured: the reading is a resident-set high-water, and under memory pressure macOS compresses the unit's pages out of the resident set as it writes them — holding 1,397 MB beside a sibling process holding 8–16 GiB, 6 of 46 plants read 610–1,217 MB whilephys_footprint, which counts compressed and swapped pages, read 1,409–1,410 MB on every one; with no sibling 15 of 15 readings were above what was held. Such a run did not register the plant, so the control cannot conclude: it now prints a SKIP naming the figure held, the figure read, the platform and, on darwin, the unit's ownphys_footprint(read by the unit only when its own resident set is below what it wrote). Everything else the control holds keeps running and keeps failing, and the judgement is held both ways on every run by forged readings through the same function — one byte under what was held is that SKIP, a peak at what was held is named — with six mutants each turning the control red by name. An incompressible plant read true in 0 of 34 runs under a sibling but lost resident pages in the heaviest of them and costs 0.8–4.9 s per run, so it is not shipped. On Linux, 63 CI runs read the plant 103–113 MB above what it held and none under-read; whether a runner compresses or swaps a child's pages is not measured.
Not in this cut: the rest of #1179 (the core-entry build at 1.46x the round-tripped one on production rigs).