pub fn capture_starvation_witness(result: &VmResult) -> Option<f64>Expand description
Witnessed-starvation evidence for capture-bearing assertions: the
host dilation D that PROVES the run was contended enough to
explain missing/degraded periodic captures, or None on a quiet
(or unwitnessed) host.
The periodic capture chain is readiness-gated (KASLR publish +
accessor-init worker) and its boundaries are wall-anchored inside
the workload window; under heavy host saturation the cold-cache
accessor build alone can outlast a short workload, so every
boundary defers past run end and periodic_fired == 0 with zero
real captures — an ENVIRONMENTAL outcome, not a capture-pipeline
regression (observed across the periodic/freeze test family on
saturated 96-CPU CI runners; the same cells pass idle). Assertions
that require captures consult this first and convert their failure
into a host-side SKIP (post_vm_skip) when starvation is proven,
mirroring the tri-state latency gates: a quiet-host miss still
FAILS (the regression the assertion exists to catch).
Witness preference: Body-phase dilation (the capture boundaries
live inside the workload window), falling back to the whole-run
schedstat D. The 1.5 bar follows PERF_ISOLATION_D_MAX’s
derivation (worst measured-quiet D ≈ 1.21; > 1.5 is unambiguous
external contention). None (no witness sampled) is NOT proof —
callers must fail in that case; absence of evidence cannot launder
a regression.