Start latency is sliced by cause — warm-pool claim, memory restore, disk-only restore, prebuild restore, inline build — because an undifferentiated p95 cannot be acted on.
Hibernation accuracy (false hibernations, missed idleness) is a first-class signal: it is simultaneously the cost metric and the trust metric.
Cost per active workspace-hour is measured on the host agent, where the host-seconds actually are, and attributed to workspace, principal and tenant.
Assumptions
OpenTelemetry collection with Prometheus for metrics and a log store for verdicts; raw telemetry 14 days, aggregated 400.
Fleet utilisation ≥ 60% of allocated host-seconds over a working day; warm pools ≤ 7% of compute spend.
Deliberately omitted
There is no correctness row at the edge and no latency row at the egress point, because neither question has an answer a developer would ever feel.
Nothing in this matrix reads the contents of a workspace; every signal is observed at the boundary.
