not done

What is not done

Lockstep is pre-beta and this page is the list. It exists because "builds and tests clean" and "works" are different claims, and a release that only advertises the first while sounding like the second is the thing worth not doing.

Nothing here is a secret or a surprise discovered late. Each item is tracked in ROADMAP.md, which is the authority; this page is the short version for someone deciding whether to try it.

Platforms

Platform	What is actually true
Linux	Built, tested and used. Every line was written here.
macOS	Compiles and passes the full test suite in CI, including the 125-row dispatch golden. No host has ever loaded the plugin. No window has been opened, no audio device opened, and the AU has never been through auval.
Windows	Compiles under MSVC and passes the vendored test suites. Its two GUI-dependent suites are excluded in CI, because a Windows runner has no interactive desktop session — so they are unrun, not passing.

macOS and Windows were compiled for the first time on 2026-10-04. Everything before that date was Linux-only. If you run Lockstep on either, a report is the most useful contribution the project can receive.

Features designed but not finished

• The recall unit (5.3). Song and Scene naming, colours, the browser and phrase copy/fork all ship. What is missing is the unit you recall a whole sound set with; "Kit" was retired as a term and the story needs re-deriving.
• Sample-playback coherence (9.23). Partially shipped.
• Per-track scale-quantize (10.6) and harmonic existing-rhythm placement (10.11).
• Cue + Scene and Cue + MIDI-out — the cue balance core shipped; these two gestures did not.
• Freeze-to-disk for record buffers.
• The Machine Module ABI (6.7) is deliberately deferred until a second consumer exists. Third-party machines are not loadable yet; the catalogue is first-party and statically linked.

Verification that has not happened

• Ear tests. Time-stretch quality, the effects A/B matrix and varispeed texture are all marked "pending ear test". They need listening, and nobody has done it since the engine changed.
• The Bungee stretch engine was bumped to v2.4.30 on 2026-10-04, which included a change to its phase maths. Stretch output will differ slightly from any previous build, and the pending ear test now applies to this version.
• auval has never been run on the AU.
• Push 1 palette has never been checked on hardware.

Known internal debts

These do not affect what you can do with Lockstep, but they are real and are written down rather than discovered:

• clang-format is advisory, not enforced. 233 of 402 files differ from the project's own .clang-format. The tree was never run through it, and reformatting half the source in one commit would destroy git blame across the codebase.
• clang-tidy is advisory. CI's whole-tree count is 14,611 findings across 73 translation units, with clang-tidy 18. (It used to report 2,227 from a glob that covered about five of them.) The tool is pinned to clang-tidy 18.1.8 so the figure measures the code rather than the runner image — a newer clang-tidy counts more because it ships more checks. Findings in a shared header count once per unit that includes it, so distinct issues are fewer than the total.
• Link-time optimisation is off. There is a circular dependency between two static archives (the engine names an editor symbol it does not own), and LTO turns it from a link that works by luck into one that fails on some toolchains. The workaround is documented in src/CMakeLists.txt; the fix is to break the cycle.
• Scene goldens are a Linux-only check. They compare pixels, and a different rasteriser moves about 1% of blocks — so chrome with no model (the inspector, transport, timeline strip, MZ rotaries, banners) is unverified on macOS and Windows.

What is solid

For balance, because a page like this is misleading on its own. The test suite is six binaries and runs in about 70 seconds. Every row of the Critical User Journey catalogue is implemented. The gesture grammar is covered by a 125-row behavioural golden that is platform-independent and passes on all three platforms, and by a modal sweep that drives all 16 modal states through a battery of interruptions. Several of the guards in this codebase fail the build rather than a test, because the bugs they prevent were invisible on screen.