The option wiring has no test seam

[xroche]

OptionsEmissionTest can only reach the mapper classes in isolation. The tables that wire a checkbox to an engine flag (fieldsSerializer id to name, fieldsDefaults, fieldsMapper) are unreachable from a plain JUnit test: they are built in OptionsMapper's static initialiser out of android.util.Pair and R.id, and the stub android.jar throws RuntimeException("Stub!") from the Pair constructor. The nested mapper classes only work because a static nested class does not trigger the outer initialiser. returnDefaultValues does not help, since the stub nulls both fields before it throws.

So a row pointing at the wrong option name, a checkbox left out of a tab's @Fields (saved but never reloaded), or a missing defaults entry all ship green. That is how the three new options landed in #77: the emitter is covered, the wiring behind it is not.

Two ways out. Add Robolectric, which gives real Pair, SparseArray and R.id without a device. Or drop android.util.Pair from the static tables for a local immutable pair, and lift the mapping loop out of buildCommandline() into a static that takes a name-to-value lookup and returns the argv. The second keeps the test dependency-free; SparseArraySerializable and the Context work stay untestable either way, which is fine because they are not where the bugs are.

Either one unlocks the assertion that matters: tick checkSitemap, get --sitemap in the argv and nothing added to the packed flag string, plus a table-consistency check that every serializer name has both a default and a mapper.

Worth noting for whoever picks this up: the engine's own self-tests (-#test=singlefile, -#test=changes) should be reachable through HTTrackLib.main, so an instrumentation test could exercise the real engine paths on a device.

[xroche]

Reviewing #78 turned up that this seam is cheaper than the issue assumes. A roughly 30-line local android/util/Pair.java on the unit-test source path makes the whole fieldsMapper table constructible, and the review's differential harness drove the real buildCommandline() that way in minutes.

Worth doing, because the table wiring is precisely what the current tests cannot see. A mutation audit of #78 killed 15 of 21 mutants, and all six survivors live in that layer, including one that remaps Sitemap to a different engine option and goes unnoticed. Two assertions would close most of it: that the table maps Sitemap, SingleFile and Changes to %m, %Z and %d, and that no emitted token contains a % followed by two letters, with the leading packed token holding no % at all.