werkator/docs/prs/2026-10-04-PR#110-nginx-cleanup-cpu-time-disks.md

WARNING: This document describes only the change applied in this PR. It may already be outdated once the next PR is merged. Historic PR-documentation is not maintained along with new PRs — treat it as a snapshot, not as current documentation.

The Problem

After PR#109 the three highest Mutation-CRAP scores of Werkator were unrelated, each small:

Method	Mutation-CRAP	Killed / eligible	Why
NginxProxyManager.cleanupStaleContainers	45.04	1 / 6	tested, but the tests passed whatever it did: the start removes the container by name once more later
UiFormats.cpuTime	43.81	1 / 11	no test of its own
MetricsConfiguration.monitoredDisks	42	0 / 9	read the configuration through Spring beans, no test reached it

The first mutation build with an active gate (PR#104) failed on an analysis error instead: two methods were INCONCLUSIVE, because a mutant had made a bounded wait endless and filled the memory of PIT's minion (MEMORY_ERROR), which werkcrappit refuses to count as killed (its ADR 0006). In NginxProxyManager.waitForPortsFree the mutant turned repeat(4) into an endless loop, and the test's sleeper did not sleep; in ConsoleBuildRunner.buildAndStream the end of the wait was asked twice, and a mutant of the first question alone made it endless and without a pause. Tonight's mutation build would have failed the same way.

The rule for a path in the instance file — ~ is the home directory, anything relative is relative to it — was written twice, for the registry's repositories and for the System page's disks.

Non-Goals

• No change in behaviour.

The Scenarios

Feature: Three small decisions, held by tests

Scenario#110.01: Processor time reads with its factor against the wall clock

So that 4:07 (2.4x) stays what the artifact page shows.

• Given processor times with and without a duration, a zero duration, nothing measured, and a negative time
• When they are formatted
• Then the factor appears only with a duration above zero, and nothing measured shows nothing

Verified by

• UiViewsUnitTest — Scenario#110.01

Scenario#110.02: The System page shows the configured disks, their paths read like every path of the instance file

So that ~, ~/… and relative paths mean the same for a disk as for a repository.

• Given no metrics.disks, and disks with ~, an absolute path, ~/… with a blank label, and a relative path
• When the monitored disks are resolved
• Then without the section the checkout's filesystem is Disk; otherwise the disks keep their order, paths resolve against the home directory, and a blank label falls back to the key
• And only ~ and ~/ expand, ~other is a name like any other

Verified by

• MonitoredDisksUnitTest — Scenario#110.02

Scenario#110.03: Only Werkator's own leftovers are removed before the managed nginx starts

So that a stale container of Werkator never blocks a port, and someone else's container is never touched.

• Given a container of the configured name, labelled nginx containers of this repository, and containers on and off the nginx ports
• When the managed nginx starts
• Then the container of the name goes first, then the labelled ones; a failed or empty listing removes nothing by label
• And on an nginx port only a container labelled as Werkator's, or named werkator-… or git-watch-origin-and-test-nginx-…, is removed

Verified by

• NginxProxyManagerUnitTest — Scenario#110.03

Scenario#110.04: A wait that a mutant makes endless fails a test instead of filling the memory

So that the mutation gate judges test quality, not memory, and the nightly build can pass.

• Given the port wait of the managed nginx and the console's wait for a build
• When PIT runs on both classes
• Then no mutant ends with MEMORY_ERROR, and no method is INCONCLUSIVE

Verified by

• NginxProxyManagerUnitTest: its sleeper fails after 100 rounds
• ConsoleBuildRunnerUnitTest
• A local PIT run on both classes: 0 MEMORY_ERROR, 0 INCONCLUSIVE

The Solution

• NginxProxyManager.ContainerInfo.isWerkatorLeftover names the rule for a stale Werkator container; removeLabelledContainers the removal by label; cleanupStaleContainers reads as the three steps.
• MetricsConfiguration.monitoredDisksOf resolves the disks without Spring; InstancePath.resolve is the one rule for a path in the instance file, used by RepositoryEntry.directory as well.
• UiFormats.cpuTime gets the test it lacked.
• ConsoleBuildRunner.awaitTerminal decides the end of the wait at one place; the test sleeper of NginxProxyManagerUnitTest fails after 100 rounds.

Measured with PIT as the nightly build runs it:

Method	Before	After
NginxProxyManager.cleanupStaleContainers	45.04	2
UiFormats.cpuTime	43.81	7.04
MetricsConfiguration.monitoredDisks	42	2
MetricsConfiguration.monitoredDisksOf	—	4
ContainerInfo.isWerkatorLeftover	—	3.03

The gate maxMutationCrap falls from 46 to 36: the highest score left is SystemMetricsCollector.sample at 35.36.

Additional Changes

• Version 1.34.7 with its release note.

Follow-up PRs

• The methods above 15 by complexity alone: SystemMetricsCollector.sample (35), BuildDefinition.applyTo (34), ForgejoWorkflowExporter.settingNotes (27), SystemMetricsView.from (24), and the ConfigLoader checks.