feat(gui): egui editor with live step grid + clickable toggles

[gustavokch]

Stage 3 — GUI port

Adds a custom nih_plug_egui editor restoring the Stage-1 openFrameworks UX (DS-UI.png): a 16-step grid with a live playhead and click-to-toggle cells, plus sliders for the latent vector, per-step pitches, timing, and tuning. The original training UI (GENERATE / Train / Make Dataset / Open Corpus / Epochs / Loss) is dropped — no runtime training, regeneration is reactive on latent change.

What changed

• src/shared.rs (new) — lock-free SharedState { steps: AtomicU16, current_step: AtomicUsize } bridging the audio thread and the editor (Relaxed; no cross-variable invariant).
• src/lib.rs — maybe_regen writes the decoded pattern into shared.steps (now the playback source of truth, replacing the old self.steps array); process reads on/off from the mask, publishes the playhead, clears it when stopped; editor() wired up.
• src/editor.rs (new) — create_egui_editor; ParamSlider widgets + a custom painter-drawn grid (fill = on, red outline = playhead, click = shared.toggle(i)); repaints each frame so the playhead animates.
• src/params.rs — persisted EguiState (600×640).
• Cargo — pin both nih_plug and nih_plug_egui to the same rev f36931f7; leaving the top-level nih_plug unpinned pulled a second newer copy and broke the Params derive.

Trade-offs

• Step toggles are runtime overrides, not params: not preset-saved, not automatable, and a latent move regenerates over them (intended generative semantics). Promoting to 16 bool params is a possible later pass.

Verification

• cargo test — 16 pass (existing sequencer/decoder + new SharedState bit tests)
• cargo build --release + cargo xtask bundle deepsteps-plugin --release — clean
• clap-validator validate — 18 passed / 0 failed / 3 skipped (unchanged)
• Manual GUI render (Carla) — not yet done, pending reviewer check

Design doc: docs/plans/2026-06-14-stage3-egui-editor-design.md

[gustavokch]

Code Review — PR #3 (feat: egui editor with live step grid)

Reviewed post-merge at the current main state. Overall: clean, well-documented work. The lock-free audio↔GUI bridge is correct and the process loop ordering is sound. No correctness bugs that hang notes or crash. Findings below are polish + stale-scaffolding, not blockers.

Overview

Adds a nih_plug_egui editor: 16-step grid with live playhead + click-to-toggle, sliders for latent/timing/tuning/pitches. Step on/off mask moves from self.steps into a lock-free Arc<SharedState> (AtomicU16 mask + AtomicUsize playhead), shared between the audio thread and editor. nih_plug/nih_plug_egui pinned to the same rev.

What's good

• SharedState design is correct: Relaxed is justified (no cross-variable invariant; toggle is an atomic RMW), well-commented, and unit-tested (mask_roundtrip_and_toggle, pack_matches_bits).
• Process loop drains pending NoteOffs before stepping, so a same-block NoteOff can't precede its NoteOn — and clamps the NoteOn offset into [0, nframes-1]. Good.
• Cargo rev-pin rationale (avoiding a second nih_plug copy breaking the Params derive) documented in the PR.

Issues

Medium — stale scaffolding in src/params.rs now masks real dead code.

• Module doc (lines 4-6): "The struct is not wired into a plugin yet (Task 13/14); dead_code is therefore expected until then." It is wired now (lib.rs builds editor::create(self.params.clone(), …)).
• #![allow(dead_code)] (line 6) is therefore no longer scoped to scaffolding — it will silently hide genuinely dead params/fields from here on. Drop it (or narrow to specific items) and let the compiler talk.
• _ParamsArc marker (lines 165-168) and the "silence the unused-Arc import" comment are obsolete — Arc<EguiState> is a live field. Remove.

Low/Medium — editor.rs repaints unconditionally.
ctx.request_repaint() runs every frame regardless of transport state, so the editor renders at full framerate (GPU/CPU spin) even when the transport is stopped and the playhead isn't moving. Gate it: only request a repaint while shared.current() != NO_STEP (transport playing), otherwise let egui idle.

Low — stale/misleading latent-range comment (params.rs:129-135).
Comment says the Stage-1 sliders were -10..10/default 0 and "Task 13 will rescale as needed." Verified the shipped reference_vectors.json latents are all in 0.028..0.981, so the chosen 0.0..1.0/default 0.5 range is correct for this decoder — no rescale is pending. Reword so a future reader doesn't think there's an open normalization task.

Low — SharedState::get/toggle are pub with no bound guard.
1 << i on the u16 mask is UB-shift / debug panic for i >= 16. All internal callers pass 0..16, but since these are public add a debug_assert!(i < 16) or document the invariant.

Low — playhead is last-write-wins within a block (lib.rs:210).
If a process block spans multiple steps, set_current(idx) only leaves the last step visible; intermediate steps are never shown. Fine at typical block sizes (≤1 step/block at 16 steps/bar), but worth a comment, or publish the playhead based on wall-clock position rather than the step loop.

Cosmetic — doc drift.

• PR body says editor is 600×640; code is 600×520 (params.rs EDITOR_HEIGHT, per commit c42a1ff). The params.rs:12 "matching the Stage-1 openFrameworks canvas" comment no longer matches either.
• editor.rs:73 comment says "4 columns" while the grid is num_columns(8) (4 label+slider pairs). Reword to "4 step columns".

Test coverage

shared.rs and decoder.rs covered; the egui editor is untested (expected — GUI). The PR notes manual Carla render is not yet done — that's the real gap. A headless render/snapshot check or at minimum a manual Carla pass should confirm the grid + playhead actually draw before this is called done.