[pumfleet]

Summary

Two pieces of infrastructure, independent of each other (happy to split them if that reviews better).

1. Tests for libs/ui (AltvaneUITests, 78 Catch2 cases)

• libs/ui/tests is a Catch2 console app built under ALTVANE_BUILD_TESTS and registered with CTest, so CI's existing ctest --test-dir Builds step runs it on Linux, macOS and Windows with no workflow change. altvane_ui is an INTERFACE library, so the app compiles it and its JUCE modules itself, under a stand-in plugin name (Altvane UI Tests).
• A small fake processor (float, skewed float, choice, bool, and a "meter" that's left out of presets) plus its Session and ParamEditor stand in for a product.
• Covered: value_format (formatting, parsing, units, signs, either minus, decimal comma, note names, junk, infinities, round trips, and a decimal-comma C locale); ParamEditor (setNow, drags, cancel, clamping and snapping, toggle, undo and redo order, undo mid-edit, gestures balanced throughout); typing into a bound Knob through its real TextEditor and the message loop, accessibility setValueAsString, and a wheel's late on_end; Session A/B (both directions, the first switch, history cleared, copy to the other slot, an out-of-range slot); UISettings; PresetManager in a temporary folder (factory presets, save and load, modified tracking, six kinds of corrupt file, saving over a name, rename collisions, case-only rename, illegal names, remove, groups, favourites, natural order, step, initialise, the JSON codec); SpscQueue, TriBuffer, FreqScale and GainScale.
• The only library API addition is PresetManager::setDirectoryOverride, which only the tests call. It keeps their presets away from the user's.

Bugs the tests found, fixed in separate commits:

• format::ms() printed 100 ms to 1 s with every decimal, because juce::String(v, 0) means "as many decimals as needed". So 250.4 ms read "250.4 ms" and 333.333 ms read "333.333 ms". This shows in the Limiter's, EQ's and De-Esser's time controls. (1bb3cc9)
• Values that round up to a range boundary had a second spelling: 99.97 Hz read "100.0 Hz", 999.7 Hz read "1000 Hz", 9.996 dB read "+10.00 dB", and 999.7 ms read "999.7 ms". (1bb3cc9)
• Under a global C++ locale with a decimal comma, which a host may set with std::locale::global, freq, gain, q and ms wrote "1,23 kHz". juce::String(double, n) imbues its stream buffer with the classic locale, but the std::ostream it writes through still uses the global one. The formatters now build the digits themselves. I found this on macOS CI and then reproduced it on Linux. (52b0bf8)
• parseNumber read text with junk before the number as 0. Typing x5 or gain 3 into any control bound with bindValue set the parameter to 0, as an undoable edit. parseFreq read 1.2.3 as 1.2 Hz. (1bb3cc9)
• PresetManager::remove cleared the favourite and the current preset before trying to move the file to the trash. When that failed, the file stayed but lost its star and the header said "Default". (b70ce8e)
• PresetManager::rename treated the preset's own file as a collision. On macOS and Windows that means "bass" could never become "Bass", and renaming to the same name failed on every platform. It now asks the file system whether the target is the same file (File::getFileIdentifier), not File::operator==. That operator ignores case even on case-sensitive macOS volumes, where a path comparison would overwrite a distinct "Bass". (b70ce8e, bdeb22a, 7fbcd0f: a case-only rename goes through a temporary name, because on Linux moveFileTo deletes the target first, and on a casefold folder the target is the same file.)

2. Scheduled CI

• Upstream Watch (new, .github/workflows/upstream_watch.yml, Mondays 06:17 UTC plus workflow_dispatch): for every products/*/UPSTREAM, scripts/upstream_behind.sh lists the upstream main commits since the pin. The workflow keeps one issue labelled upstream-watch with a table and per-product commit lists. It creates the issue when something is behind, rewrites the body each week, comments only when the counts change, and closes it once everything is current. Permissions are contents: read and issues: write. It doesn't touch code or open PRs. Commit subjects go in code spans so upstream's @mentions and #numbers don't ping or link anything here.

• Current lag (2026-11-02):

  Product	Upstream	Pinned at	Behind
  compressor	ZL-Audio/ZLCompressor	2f7b45c	1 commit
  eq	ZL-Audio/ZLEqualizer	3ce598e	current
  limiter	ZL-Audio/ZLLimiter	412c844	24 commits (22 at the last check)

• Sanitizer (Saturdays) and Heavy Test (Sundays), both at 02:23 UTC, keep workflow_dispatch and its product input.

  • Run on a schedule, both workflows would have had an empty product, so they would have configured with -DALTVANE_PRODUCTS= (nothing built) and failed on a missing .env.
  • A small products job now picks the input's product, or every product on the schedule. build_and_test runs product × runner. On the schedule it skips runners a product's PLATFORMS file rules out (Dialogue on Linux), the same way the top-level CMakeLists does. A dispatched single product still builds on every runner, as before. Excluding every combination for a runner would turn its include entry into a job with no product. The existing include entries are untouched.
  • The concurrency group now includes the workflow name. With ${{ github.ref }} alone, the scheduled runs on main, pushes to main and the other scheduled workflow would cancel one another.
  • What they build and test is unchanged.
  • The weekly cost: 22 sanitizer jobs (12 on macOS) and 17 heavy-test jobs (6 on macOS, 6 on Windows). The repo is public, so these minutes are free.

Testing

• Linux dev box (clang-21, Release, VST3, -DALTVANE_BUILD_TESTS=ON, De-Esser plus the UI tests): xvfb-run ctest -j6 passes 128 of 128. The UI tests are 78 cases and 1151 assertions. The locale test sets its own comma numpunct locale, so it runs everywhere, and it fails on Linux without 52b0bf8.
• Each bug-fix commit was checked in order. The tests-only commit passes on the unfixed library. Each fix commit's new tests failed on the code before it, except the remove happy-path test. The case-only rename can't fail on Linux's case-sensitive file system, so CI's macOS and Windows jobs are what check it.
• The full build of all six products and pluginval are left to CI.
• actionlint 1.7.7 passes on all workflows except one existing error: sanitizer.yml's unreachable Windows step uses matrix.vcvars64_path, which isn't defined. I left it alone, since another change is editing those files.
• codex exec review --base origin/main, four runs:
  • Run 1 found the rename identity check and the matrix job with no product when dispatching Dialogue. Both are fixed in bdeb22a.
  • Run 2 was clean.
  • Run 3, after the CI fixes, found that a case-only rename on a casefold Linux folder would delete the preset. That's fixed in 7fbcd0f.
  • Run 4 was clean.
• The first CI run failed on Windows, where clang-cl couldn't compile JUCE's bundled libwebp. The test app now uses the same JUCE settings that altvane_add_plugin gives products, including JUCE_USE_WEBP=0 (7a3ce25). The macOS run found the locale bug above.
• Upstream Watch can't be dispatched from this branch: gh workflow run --ref returns 404 until the file is on main. So I ran its report step locally in a shallow clone, as actions/checkout makes one; it produced the table above. I ran its issue step against a stand-in gh in five situations: behind with no issue, unchanged counts (edit only), changed counts (edit and comment), current with an issue open (close with a comment), and current with no issue (nothing).
• I ran the products job's script locally for both workflows, scheduled and with dialogue and eq inputs.

Review

• The tests define JucePlugin_Name, PRODUCT_NAME_WITHOUT_VERSION and VERSION themselves. If altvane_ui starts reading another plugin macro, this target needs it too.
• On Linux, the preset tests keep their folders under ~/.cache rather than /tmp. gio only trashes files on the home folder's file system, and /tmp is often tmpfs. They also point XDG_DATA_HOME at a trash of their own, so nothing lands in the user's trash. On macOS and Windows, the remove test uses the real Trash or Recycle Bin, with a file named Doomed.json.
• Not fixed here: JUCE's serialiseDouble, which saves state and XML attributes, goes through the same stream. Under a comma locale, saved state and presets would hold "0,5"-style values. That belongs in the saved-state work or upstream in JUCE.
• parseNumber still ignores what follows the number, because products rely on that (parseMs reads "20 ms"). It still treats a comma as a decimal point, so "1,000" reads as 1. And -inf doesn't parse. The tests pin all three behaviours.
• The runner names in sanitizer.yml and cmake_heavy_test.yml are now listed twice: in the include entries and in the products job's RUNNERS. A new entry needs both.
• Scheduled workflows run only from main. GitHub disables them after 60 days without repository activity.

🤖 Generated with Claude Code