fix: [ANDROAPP-7775] report the project's real test coverage

[andresmr]

Coverage was reported for unit tests only, over roughly a fifth of the codebase. Two
independent causes, both silent — a missing coverage report is a warning, not an error,
and CI hid even that (if-no-files-found: warn, continue-on-error: true).

Cause 1 — instrumented coverage never left BrowserStack

The UI suites run on BrowserStack's devices, so their JaCoCo execution data was written
there and nothing brought it back. code-quality also ran alongside those jobs rather
than after them, so it could not have used the data even if it had arrived.

Cause 2 — the exclusion list removed 82.3% of the code

jacoco/jacoco.gradle.kts excluded 1803 of 2191 compiled classes in app. What the
report called the project's coverage was the coverage of the 18% that survived.

Measured per pattern on app's dhis2Debug output:

classes removed	pattern	what it actually removed
1046	**/*$*	every nested class, lambda and coroutine state machine
664	**/*$*$*.*
332	**/*Module*.*	anything with "Module" in the name, not just DI modules
197	**/*Activity*.*
153	**/databinding/*.*	(legitimately generated)
127	**/ui/*.class	including plain mappers that happen to live under ui/
110	**/*Dialog*.*

Among the casualties of **/*$* were 197 named nested types carrying real logic:
sealed subclasses (ConvertTaskResult$Message), nested enums (OpenIdSession$LogOutReason),
$DefaultImpls — the actual bodies of interface default methods — and builders.

What changed

1. Instrument debug builds — enableAndroidTestCoverage = true on the debug build
   type of app and form, so instrumented runs emit .ec data. Debug only; release
   builds and the shipped APK are unaffected.
2. Retrieve the data — new scripts/fetchBrowserstackCoverage.sh reads
   .devices[].sessions[].id from GET /app-automate/espresso/v2/builds/<id> and
   downloads each session from .../sessions/<session>/coverage. "coverage": true
   added to the three build requests. The script never fails its caller — missing
   coverage should not turn a green test run red — and runs before the pass/fail exit so
   a failed suite still contributes what it covered.
3. Merge before analysis — the BrowserStack jobs upload what they retrieved,
   unit-tests also uploads raw **/build/jacoco/*.exec, and code-quality now waits
   on all of them, copies the .ec files into each reporting module and regenerates the
   XML before Sonar runs.
4. Exclude only generated code — the list now keeps R/BR/BuildConfig/Manifest, Data
   Binding, Dagger/Hilt, other annotation processors, $WhenMappings, $$serializer,
   ComposableSingletons and test classes, and nothing else. Android UI classes are no
   longer excluded on purpose: the instrumented suite exercises them, and hiding them
   understated the very tests that cover them.
5. Unpin JaCoCo — required, not incidental. See below.
6. verify-dependency-metadata.yml gets a concurrency block — see below.

Required: unpinning JaCoCo

Turning on instrumentation broke the build outright:

Execution failed for task ':form:mergeExtDexDebugAndroidTest'
  > Failed to transform mockito-kotlin-6.1.0.jar ... useJacocoTransformInstrumentation=true
    > Execution failed for JacocoTransform:
      .../byte-buddy/1.17.7/byte-buddy-1.17.7.jar

enableAndroidTestCoverage makes AGP run JacocoTransform over every dependency on
the androidTest runtime classpath. The root build pinned all org.jacoco artifacts to
0.8.10 through a resolutionStrategy, overriding whatever version AGP asked for, and
0.8.10 cannot instrument byte-buddy 1.17.7's bytecode. The pin predates instrumented
coverage and cannot survive it.

The pin is removed and the jacoco plugin's toolVersion is wired to
libs.versions.jacoco, so the catalog governs the report engine rather than Gradle's
bundled default. Modules that instrument (app, form) use AGP's own jacoco version for
the transform — the version AGP is tested against.

Verified locally that :form:mergeExtDexDebugAndroidTest, the task that failed, builds.

gradle/verification-metadata.xml gains the entries the newer artifacts require.

Result — verified on CI, every job green

instrumented coverage files retrieved: 3

job	.ec retrieved
run-ui-landscape	1.3 MB
run-ui-portrait	1.3 MB

	coverage	lines measured	lines covered
develop	10.6%	47,579	5,486
this branch	17.9%	61,224	12,298

Read the counts, not just the percentage. This measures 29% more code and credits
2.2× more covered lines than develop. Roughly 6,800 more lines of genuinely tested code are credited than on develop, and a
large body of untested UI code that was hidden from the denominator is now measured.

Exclusions: 82.3% → 25.4%. Classes analysed in app: 329 → 1074.

Also fixed: Verify Dependency Metadata pushed empty commits

That workflow had no concurrency block, so its runs stacked up. Each compares its
regenerated metadata against the commit it checked out rather than the branch tip, so the
slower of two overlapping runs still saw a difference after the faster one had already
committed the same fix, and pushed a commit with a byte-identical tree:

08:06:56  run A starts at ea6d2506c
08:23:47  run B starts at e2d3f3469
08:41:02  run A pushes 397a0a2da   (+18 lines -- legitimate)
08:46:10  run B pushes 04caca0c7   (identical tree -- empty)

Every such push cancelled the in-flight CI run. Four runs died this way while this work
was being verified. Serialising per ref means only the newest run survives, and it always
regenerates from the newest tip.

Note on the form suite

ANDROAPP-7748 (#5034) retired the form instrumentation CI job while this PR was open, and
this branch adopts that fully: the BrowserStack form script is deleted, form is out of
the coverage staging list, and form/build.gradle.kts is back to develop's version —
develop has no form/src/androidTest sources left, so instrumenting that module would
cost build time for data that cannot exist.

Instrumented coverage therefore comes from two device jobs rather than three. That is why
this figure moved from 19.1% when first measured to 17.9%: the denominator is unchanged at
61,224 — only device-sourced covered lines fell. The form module's replacement JVM tests
are counted through the normal unit-test path.

Trade-off

code-quality now waits on the device suites, so the Sonar result arrives later than it
used to. That is inherent to including their coverage.

Not in this PR

• Making the metadata check compare against origin/<branch> rather than its own
  checkout. The concurrency block stops runs from overlapping, but the comparison is
  still against the wrong baseline if anything else pushes mid-run.
• Deleting the stale SonarCloud required status check from the rulesets — nothing
  reports it, so it can never pass. Needs an admin. (SonarCloud Code Analysis and
  code-quality are both legitimate and should stay: one is the scan, the other the
  Quality Gate verdict.)

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud