Pokered Save Editor 2: versioning, i18n, a Linux test container, and real GUI tests
A heavy infrastructure day — single-source versioning, translation plumbing, a Dockerised Linux harness, and tests that drive the live app.
A lot of foundation went into Pokered Save Editor 2 this day. None of it is glamorous alone, but together it’s the scaffolding the rest of the project leans on.
One source of truth for the version
The version number now lives in a single VERSION file, which CMake reads and
propagates into the runtime, the About screen, and the Windows executable resource —
with the git commit appended as build metadata. There is no longer a hardcoded version
anywhere to drift:
file(READ "${CMAKE_SOURCE_DIR}/VERSION" PSE_VERSION)
string(STRIP "${PSE_VERSION}" PSE_VERSION)
project(PokeredSaveEditor VERSION ${PSE_VERSION} LANGUAGES CXX)
# → pse_version.h: "0.7.0-alpha+g<hash>"
Internationalisation
UI strings are now wrapped for translation, with per-locale catalogs compiled and
embedded and a QTranslator installed at startup (qsTr("…") in QML, tr("…") in
C++). The source language is US English; game-data names are deliberately out of scope.
Actual language switching is deferred until there’s a second locale and an options
screen — the plumbing is what landed here.
A Linux test container
A Dockerised build-and-test environment mirrors the dev kit and adds what the Windows kit can’t run — Address/UB sanitizers — plus a real virtual display and coverage:
./docker/dtest.ps1 standard # offscreen ctest
./docker/dtest.ps1 asan # AddressSanitizer + UBSan (not available on the Win kit)
./docker/dtest.ps1 xvfb # real virtual X server
./docker/dtest.ps1 coverage # llvm-cov
On its first run all four variants passed, sanitizers clean, line coverage just under 90%.
Tests that drive the real app
Two new kinds of test went in. One loads every screen through the real engine and fails on any load error or warning — the class of problem the C++ unit tests can’t see because they never instantiate the UI, and exactly what had let a non-opening Credits screen slip through earlier. The other boots the actual application headless and drives it: navigate every screen, edit fields through the live UI, save, reopen, assert the changes persisted — including a byte-fidelity test that browses everything without editing and confirms not a single save byte changed.
$ ctest --output-on-failure
Start 1: tst_roundtrip
...
100% tests passed, 0 tests failed out of 66
Those tests immediately caught real crashes — a garbage save file crashing on load, the shop screen aborting when opened with an empty cart, and a badge-count function stubbed to always return eight. The day closed with an end-to-end redesign of the Credits/About screen onto a data-driven back end, so adding a credit is now a single data edit.