feat: v0.2.0 deep-dive rework, M2 preset system, i18n frame

[yves-vogl]

Summary

Research-derived deep-dive rework of Triptych's per-band compressor voicing, plus the suite-wide M2 preset system and a German i18n frame. Ships as v0.2.0.

Deep-dive voicing rework (docs/design-brief.md, sourced in docs/research-notes.md)

• New soft-knee stage. v0.1 wrapped juce::dsp::Compressor directly, whose gain formula (gain = (env < threshold) ? 1.0 : pow(env * thresholdInverse, ratioInverse - 1.0)) is a hard knee with zero transition width - confirmed by reading the JUCE 8.0.14 source, and no knee parameter existed anywhere in v0.1's ParameterLayout.cpp. BandCompressor now drives juce::dsp::BallisticsFilter directly (the same envelope-follower class Compressor used internally) through a from-scratch gain computer, src/dsp/KneeGainComputer.{h,cpp}, implementing the standard quadratic soft-knee interpolation (Giannoulis/Massberg/Reiss, JAES 2012) with a new per-band Knee parameter (0-100%, default 50%). Knee's extent is threshold-relative: at 100% it spans [2 * thresholdDb, 0], matching the Weiss DS1-MK3 manual's documented "0 to twice the threshold value" convention. At 0%, KneeGainComputer::computeGainLinear() takes a dedicated linear-domain fast path that reproduces v0.1's exact hard-knee formula bit-for-bit - a tested regression guarantee (tests/KneeGainComputerTests.cpp, tests/BandCompressorTests.cpp).
• Per-band default recalibration. v0.1 gave Low/Mid/High one identical, uniform default (threshold -18 dB, ratio 4:1, attack 10 ms, release 100 ms). Research into the mastering multiband-compressor reference class documented two opposite philosophies (peak control vs. density/knit-together) and band-position-dependent ballistics that v0.1's uniform default matched neither of. New defaults: Low -24 dB/2.5:1/25 ms/180 ms, Mid -30 dB/1.8:1/10 ms/100 ms (the v0.1 anchor), High -20 dB/2:1/5 ms/55 ms - encoded as a standing invariant (lowAttack > midAttack > highAttack, lowRelease > midRelease > highRelease), not just a one-time value, in tests/VoicingGuaranteesTests.cpp.
• Explicitly out of scope (named, not implied as parity with the reference class): external sidechain and adjustable crossover slopes (already deferred from M1, unchanged reasoning); newly identified as v0.2.0 gaps and deferred to M2+/M3: per-band Mid/Side processing, an RMS/Peak-selectable detector mode, and program-dependent/auto-release. See docs/architecture.md's "Deferred from v0.2.0" section and docs/design-brief.md's honesty section.
• This voicing is research-derived, sourced from published manufacturer manuals (Weiss DS1-MK3, FabFilter Pro-MB) and a mastering-engineer technique article (Sound on Sound), not measured against reference hardware - flagged explicitly in the manual/CHANGELOG per the sibling Miserere v2 brief's framing.

M2 preset system (.scaffold/specs/preset-system-m2.md)

• src/presets/ (PresetManager/PresetBar/Localisation) copied verbatim from the basilica-audio/nave pilot implementation - see docs/preset-system-notes.md in that repo for the replication recipe.
• Eight factory presets (presets/factory/*.json, docs/presets.md): Default, Density Glue, Peak Control, Low-End Tighten, De-Harsh Highs, Mastering Safety Ceiling, Parallel-Style Density, Hard Limiter Ceiling.
• German localisation of the preset bar's frame strings (resources/i18n/de.txt), selected automatically for de* system languages; parameter names/units are never translated.
• State migration is tolerant: a pre-v0.2.0 session missing the new Knee parameter IDs loads cleanly with Knee at its declared default (50%), not 0/garbage - relies on AudioProcessorValueTreeState::replaceState()'s own existing "missing ID keeps current value" behaviour, no special-case code needed (tests/StateTests.cpp's migration-tolerance test).

Housekeeping

• ICON_BIG "docs/assets/icon.png" was already present in CMakeLists.txt (added by the concurrent patch-release PR) - verified, not re-added.
• .github/workflows/release.yml's create-release fix was already merged to main before this branch was cut - verified present.
• docs/manual.md updated for the new Knee parameter, per-band defaults, presets, and i18n.

Test plan

• cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Debug (CPM_SOURCE_CACHE, native arm64) configures cleanly.
• cmake --build build --target Tests Triptych_Standalone --parallel 4 builds cleanly (no new warnings beyond JUCE's own pre-existing splash-screen pragma note).
• ctest --test-dir build --output-on-failure - 73/73 tests green (up from 42; all v0.1 test intent preserved, two tests pinned to Knee 0% to keep their exact v0.1 regression assertions, adjusted defaults elsewhere per ParameterTests.cpp).
• CI matrix (macOS + Windows, pluginval strictness 10 + auval) - pending this PR's checks.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com