92.8% line (2183/2352), 81.3% branch (1091/1342), 584 JVM tests in 87 classes, measured
2026-09-02 on the tree this branch creates rather than quoted from a PR body.
The shape of the wave is worth more than the number, and it is different from the two
before it. Waves 1 and 2 were finding uncovered code; by wave 3 there was little of that
left, so the gaps had to be sorted before any test was written. Coverage gaps -- filtered
to sites where JaCoCo reports mi > 0, which is what separates a real gap from a partial
branch on a compound condition, and which cut the candidate list roughly in half. And
assertion gaps, where JaCoCo is green and nothing checks the answer: MainActivity's rail
and bottom bar were both executed and transposing them passed the entire suite, as did
swapping the two progress-notification strings and swapping Content's two destinations.
No coverage number would have found any of the three.
Naming the required mutation per ticket earned its keep three times, each recorded with
what the weak assertion actually was. Also recorded: a green mutation is only evidence
when the mutation is a real change -- one classify reordering was semantically equivalent
for every reachable input, and a bad mutation and a weak test look identical in the output.
Two entries came back as not gaps, which is a result rather than a shortfall:
ContainerCapabilities:282's exclude filter cannot drop anything, and probeForConcat's
catch arm is unreachable on this runtime -- Robolectric's MediaExtractor never throws from
setDataSource, measured across four input shapes.
Both denominators moved, in opposite directions and for different reasons, so they are
stated rather than folded into the percentage: 1340 -> 1342 branches from MediaProbe.merge,
2348 -> 2352 lines from the ConcatJoiner interface. Neither is new untested code.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Recovered onto main after hitting #160's trap for real. #189 was opened against
test/concat-engine-seam and, unlike #184-#188, never retargeted to main before merging --
so it merged into a branch that had already been merged and left behind. GitHub reported
`merged`, the PR shows MERGED, and none of it was on main: `git merge-base --is-ancestor`
is what said so, one line, immediately.
That check is the entire reason this was a five-minute recovery rather than a coverage
entry that silently stayed three points stale. The failure mode is exactly what CLAUDE.md
warns about; what it did not say, and now would, is that the auto-retarget it describes
belongs to GitHub's stacking feature, so a stack opened with plain `gh pr create --base`
has to be retargeted by hand for every single PR -- and missing one is invisible until you
check ancestry.
Numbers re-measured on this tree with --rerun-tasks rather than inherited from the branch
they were taken on: 584 tests, 0 failures, 2183/2352 line, 1091/1342 branch. Same figures,
earned again.
JMR-dev