[pumfleet]

Summary

Moves code that was copied between products into the shared libraries, with no change to behaviour, sound or pixels.

Moved

• Test helpers → libs/testing (new INTERFACE target altvane_testing). alloc_counter.{hpp,cpp} and juce_init.cpp were copied into all six products' tests/ (not only Dialogue and the Leveler). altvane_add_tests now links altvane_testing into every test binary, and tests include <altvane_testing/alloc_counter.hpp>.
  • The copies behaved identically. The De-Esser's and Dialogue's alloc_counter.cpp only restyled the Windows aligned-allocation branch (if (void* p = …), #ifdef). The shared copy is the EQ's and Compressor's (the newest), unchanged. Nothing got stricter or looser.
  • It is an INTERFACE library, like altvane_ui. The listener's static registration and the operator new replacement must be compiled into each test binary, and a static library would drop them because nothing refers to them. JUCE comes from the plugin the tests link (altvane_link_plugin_code), so it isn't linked again.
• Peak exchange → ui::PeakExchange (libs/ui/util/peak_exchange.hpp, header-only, next to TriBuffer and SpscQueue). It replaces storePeak/takePeaks in the Leveler, De-Esser and Dialogue processors. The memory orders are the same: a relaxed CAS loop to raise the peak, then a relaxed exchange with 0 converted to dB with a −120 floor. It is lock-free (static_assert) and doesn't allocate. takeInputPeaks/takeOutputPeaks keep their signatures.
• OutputRail → ui::OutputRail (libs/ui/widgets/output_rail) for Dialogue and the Leveler, whose rails were the same widget: the same layout, "OUT" caption, shared colour tokens and caption font, and a click that resets the peak hold. It takes the processor and a function that takes the output peaks. To show a mono output in both bars, Dialogue checked the main bus and the Leveler checked all output channels. The Leveler has a single output bus, so both checks give the same answer, and the shared rail checks the main bus. No colour parameters were needed.
• dB text → ui::format:
  • decibels(db, force_sign) shows one decimal with the typographic minus, an optional plus, and no sign within 0.04 dB of zero. It replaces the Leveler's and De-Esser's formatDb and the formula inside Dialogue's formatLevel.
  • parseLenient(text) maps the typographic minus to a hyphen, then keeps only digits, signs and points. It replaces the parsers in the Leveler's and De-Esser's params, Dialogue's level parameters and the Leveler's rider fader.
  • The products keep thin wrappers (params::formatDb, the local parseNumber), so the product still decides whether gains are signed.

Left alone, and why

• EQ's OutputRail. It isn't the same widget: it has a display-range button and menu in place of the caption, a 22 px top row instead of 16, peaks from the SpectrumWorker, and a repaint on global changes. Sharing it would mean hooks for the header, the hit-testing and the repaint condition, which is more code than the copy. The De-Esser's, Limiter's and Compressor's rails show gain reduction and are different widgets.

• History views. The parts they could share are already shared. The Leveler draws through ui::HistoryRing, TimeAxis, LevelAxis and traceHistory from libs/ui/widgets/history. The De-Esser keeps its own:

  • a frame buffer, grouped on a fixed grid so the picture doesn't shimmer;
  • a cached grid image, span chips and freeze;
  • a relative threshold drag with the host's context menu.

  Their hover readouts also differ in geometry, rows and styling. A shared base would hold almost nothing, so I didn't add one.

• Formatters that round differently. The Compressor's formats::db and the Limiter's level take the sign from the rounded value: −0.045 shows "0.0", not "−0.0". The Compressor's also drops the decimal from 100 dB. ui::format::gain (the EQ's) shows two decimals below 10 dB. Merging any of them would change displayed text. ui::format::parseNumber (used by the EQ, Compressor and Limiter) also differs from parseLenient: it reads a comma as a decimal point, stops at letters, and returns nothing without digits. So it stays as it is.

• Dialogue's sensitivity and the Leveler's percent-knob parsers. They don't map the typographic minus, so "−50%" reads as 50 there. Switching them to parseLenient would change that, so they are untouched. That is probably an oversight and worth a separate fix.

• The EQ's SpectrumWorker output peaks. These use a similar CAS loop, but with sequentially consistent ordering and one channel at a time. That is a different contract, so I didn't fold it in.

• K-weighting (Limiter only), per the brief.

• captionFont() is still defined in a few places (the Leveler's tokens, Dialogue's chrome, the Limiter's common header, and an anonymous one in libs/ui). The new rail uses its own anonymous copy, as widgets.cpp does. Making it public API was outside this change.

No state save/load code, .github/workflows or vendored engine directories (dsp, zlp, chore) are touched.

Testing

Local, on the Linux dev box (clang-21, Release, -DALTVANE_BUILD_TESTS=ON -DALTVANE_BUILD_TOOLS=ON -DALTVANE_LTO=OFF -DALTVANE_FORMATS=VST3 -DALTVANE_SIMD_TARGET=AVX2 -DALTVANE_EQ_BAND_NUM=8). Dialogue was built with a local, uncommitted Linux-x86_64 line in its PLATFORMS file.

• Full build: every plug-in, test binary and tool, with no new warnings.
• ctest (xvfb-run), all six products including Dialogue: 229/230 pass. The exception is the models load under a locale with a decimal comma, which skips itself because the box has no decimal-comma locale (only C, POSIX and en_US), and ctest counts an all-skipped run as failed. Neither the test nor the model loader changes here, and it passes on macOS and Windows CI.
• The allocation tests aren't vacuous: nm shows all six test binaries define the replacement operator new, alloc_counter::begin and JuceRunListener from libs/testing.
• Snapshots, before and after: I built each tools/snapshot on unmodified main and on this branch, then rendered the Leveler (7), Dialogue (8) and De-Esser (9) snapshots at scale 2. All 24 PNGs are byte-identical to main's. Two baseline runs were also byte-identical to each other, so the renders are deterministic and a byte comparison is meaningful. The EQ, Compressor and Limiter editors aren't touched; their only change is in their tests.
• Formatting equivalence: a scratch program compared ui::format::decibels and parseLenient with the code they replace (the Leveler's, De-Esser's and Dialogue's formatters and parsers). It found 0 mismatches over 23M floats and 200k typed strings. The floats were every value within ±0.035–0.045 and 64 ulps either side of each 0.05 rounding step up to ±200 dB, plus a log sweep from 1e-30 to 1e6. The strings were random mixes of digits, signs, points, commas, spaces, units and the typographic minus, plus edge cases. The only possible difference is text without any digits: it now always parses as +0 where it could have been −0 before, and the two compare equal.

CI, run 41495002417, all green:

• macOS: ctest 230/230, pluginval for all six products, and auval validated every AU.
• Windows: ctest 230/230 and pluginval.
• Linux: ctest 188/188 (no Dialogue) and pluginval.
• Vendored engines are unmodified: passed.

Not verified: nothing was run in a DAW. Real-time timing isn't affected (the audio-thread code is the same CAS loop and exchange, moved into a header) and wasn't measured, since this box's numbers aren't meaningful.

Review

• codex exec review --base origin/main, round 1: "No actionable regressions were identified." There have been no code changes since, so no further rounds.

🤖 Generated with Claude Code