[codeskine]

What changed

A technical review of the plugin (DSP, real-time, VST3/AU, build), the fixes it recommended, and the measurements a sound designer would want next. The report and the experiments it cites are in review/; the build and test logs were kept out of the repository.

Fixes from the review (349127c)

• Echo start-up (D-02): the echo's delay line started at 5 s and glided towards the knob for seconds, settling up to 3 ms short of it, so the first repeats after loading (or after a NaN recovery) came late and warped. It now jumps to the knob on the first tick. New test echoStartsAtKnobTime.
• Bypass latency (D-04): processBlockBypassed delays the input by the declared latency, so hosts at rates other than 48 kHz keep the bypassed audio aligned. New ui_test checks.
• Parameter registry (D-03): ParameterRegistry keeps its parameters in a fixed array and runs the handlers in an explicit order, the one libc++'s unordered_map happened to give, instead of depending on the standard library's iteration order.
• Noise generator UB (D-01): DaisySP's WhiteNoise overflowed a signed int32_t on every sample, which stopped ASan/UBSan at the first sample. It now multiplies as uint32_t: same bits. Recorded in third_party/DaisySP/PATCHES.txt.
• Sample rate 0 (D-06): prepareToPlay clamps a placeholder rate of 0 to 48 kHz.
• AU in CI (D-05): CI builds Drosera_AU, installs it for the runner and runs auval and pluginval on it. Passed on its first run.
• JUCE pin (D-10): fetched by the commit hash of 8.0.9, not by the tag.

Measurements and the fix they led to (ec6e150, 0237f49, bac7cf6)

• The Standalone's output, recorded through BlackHole, matches the engine rendering the same state (levels, partials, echo period, stereo).
• review/2026-11-03-sound-design-measurements.md: preset levels and birth times, the MIDI pitch map with and without reverb, repeatability of a loaded state, mono compatibility, ten-minute runs at the extremes, extreme settings, and the same presets at 44.1/48/96 kHz.
• Repeatable loading: a host hands the plugin its state after prepareToPlay, so the parameters glided from the defaults and five presets settled on a different partial (up to an octave away) depending on when the state arrived. The block after setStateInformation now takes the parameters at once; presets chosen in the window still glide. New ui_test check.
• Manual: the MIDI tab and the known limits give the measured detuning from Reverb Mix (dry, notes 18–84 within 17 cents; at 35 % Mix, notes below 48 up to 46 cents off).

Reference renders

Rule 2 of CLAUDE.md: the echo fix changes the three reference renders with echo (p05, p06, p10) throughout, because the old one-pole never reached the knob; the seven without echo are byte-identical, which also shows the registry change alters nothing. SHA256SUMS is regenerated. tests/fixtures/state-baseline.bin is untouched.

Verification

• voice_allocator_test, render_test (472 checks), ui_test, plugin_host_test with the baseline fixture, check_reference.sh: all pass on every commit.
• The three suites pass under ASan + UBSan with zero reports at 349127c (before that change UBSan halted on the noise generator). Not rerun for the later commits, which add a flag, documents and tests.
• Not run locally: pluginval and auval (CI does), Intel build.

For the reviewer

• The echo's one-pole still settles slightly short of the knob after a knob move (inaudible, noted under D-02); snapping it would change the checksums again, so it is left as is.
• Still for the ear: a listening pass over the 16 presets with the measurement tables at hand, MIDI phrasing, knob movement, real sources into the Resonator, the PONG's aliasing at full Drive.