Minimal dependencies, on purpose

This page argues that minimizing dependencies is a reliability decision rather than a preference, and that the test suite is shaped by the same constraint.

Single-host mode has no required installed npm package, build step, or framework. This machine's job is running CI; a monitoring daemon that breaks unattended because a transitive dependency changed is worse than no daemon.

The charts are HTML and CSS rather than SVG, which for horizontal bars is the better tool and not a compromise: percentage widths are responsive for free, and text stays at its real size instead of being scaled by a viewBox — which is how SVG charts end up with 9px axis labels on one screen and 20px on another.

There is one isolated exception:

• Optional PostgreSQL HA dynamically loads the locked pg driver. A normal single-host start never imports it, and fleetctl.sh install installs it only when a database URL file is configured.

Tests

cd dashboard && npm test          # Node's built-in runner, no dependencies
../scripts/test-drain.sh          # drain and resume, against a fake fleet
../scripts/test-ephemeral.sh      # the ephemeral reaper's refusals

Everything runs against fixtures in /tmp with no network and no fleet, so it is safe on the machine that is also running CI.

The decision-making modules are pure functions for this reason: queue-cause, autoscale, placement, simulator and forecast take plain data and return a decision plus its reasoning, executing nothing. The simulator's tests are hand-worked timelines — three ten-minute jobs arriving together serialize to waits of 0, 10 and 20 minutes on one runner — so the expected numbers can be checked by reading them rather than by trusting the code that produced them.

The shell tests are mostly about what the scripts refuse to do, since those are the paths that cost something: a busy runner is not interrupted by a drain, health.sh --repair does not revive a drained runner, and the reaper never deletes a directory with a live process in it.