docs: refresh README · ARCHITECTURE · FEATURE_ROADMAP · CHANGELOG for v2.0.x

README:
- Main mermaid diagram redrawn — adds SessionLiveSnapshot,
  TakeLaneSnapshot, ModSnapshot, RTVoiceScheduler, AUv3ViewCache,
  comping canvas, Help tab, the priority chain (session → arrangement
  → order list), audio + MIDI clip dispatch, take comp triggers,
  metronome shared trigger, and the Tools layer fan-out.
- New Tools + Collaboration data flow mermaid — shows how every
  offline actor (Stem · MIDI · Masking · DAW-Project · Collab) hops
  through Timeline.publishSnapshot to the engine, and where each
  writes (disk · arrangement · EQ · wire format).
- Module breakdown updated for every module.
- New feature sections: Recording + Comping, Modulation, Tools, Open
  Interchange, Collaboration Foundation, Help & Onboarding.
- UAT count: 270+ → 299.

docs/ARCHITECTURE.md:
- Module-graph mermaid annotates every module with today's additions.
- Module-role table refreshed.
- Render-pipeline diagram shows priority chain + automation +
  modulation + audio/MIDI dispatch + take comp + metronome as
  parallel branches converging on activeChannelIndices, plus the RT
  scheduler as an optional parallel path.
- Snapshot-lifecycle diagram: third atomic (_modPtr) + third dealloc
  queue (modDealloc).
- "Arrangement consumption" → "Playback priority chain" — diagram
  redrawn around the three handles (session live, arrangement, take
  lanes) all converging on the voice trigger.

docs/FEATURE_ROADMAP.md:
- Pie chart: 39 → 47 items, refreshed buckets.
- Today / Next / Later flowchart rewritten around 299 / 0 +
  Demucs/Magenta conversion + automerge-swift / SwiftGit2 vendoring
  as the next-release gates.
- Shipped list: 16 new bullet entries covering everything from
  today's session.

CHANGELOG.md:
- [Unreleased]: full breakdown of session live, take comping +
  region playback, modulation DAG, MaskingAnalyzer, StemSeparator
  scaffold, MIDIGenerator scaffold, comping UI, AUv3 native UI
  embedding, CollabSession foundation, DAW-Project exporter, plus
  the audio + MIDI clip render-path consumption from the prior
  v2.0.1-followup work.

Author: mstits

CHANGELOG.md

@@ -6,6 +6,102 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+### Added
+- **Session-view live-cell render path** — `SongSnapshot.sessionLive`
+  carries a render-thread-safe `[SessionLiveSnapshot.Cell]` derived from
+  `SessionGrid.livePlayback` + per-column `liveSinceBeat`. The render path
+  treats live cells as priority-1 overrides over arrangement clips and the
+  order list (matches the Ableton "session overrides arrangement" rule).
+  Pattern cells substitute the row; `.audio` and `.midi` cells loop with
+  per-loop trigger / event dispatch via shared dispatch helper. Quantized
+  launches now actually advance: `Timeline.publishSnapshot` calls
+  `SessionGrid.advanceLaunches(nowBeat:)` every UI tick. Plus
+  `PlaybackState.arrangement` and `PlaybackState.sessionGrid` were added
+  so the live engine consumes them — pre-fix only the test harness did.
+- **Take comp regions + render path playback** — `TakeLane.regions:
+  [CompRegion]` for region-based active-take selection.
+  `TakeLane.cacheInBank(_:)` allocates each take's interleaved PCM into the
+  unified sample bank; the render path triggers a voice from the cached
+  bank region whenever the playhead crosses a region's `startBeat`.
+  `RecordingTake` gained `bankOffset`/`bankLength` (runtime-only — never
+  persisted). `TakeLaneSnapshot` is the RT-safe handoff.
+- **Modulation routing DAG** — new `ToooT_Core/Modulation.swift` with
+  `ModSource`, `ModSlot`, `ModSnapshot`. Sources: LFO, sidechain peak,
+  envelope, MIDI CC. Destinations: channel volume / pan / send, bus
+  volume, master volume. `RenderResources.modBuffer` is a flat 128-slot
+  Float* allocated at engine init. `evaluateModulation` runs at every row
+  tick after automation, additively writing source × amount into the same
+  destination fields automation already touches. Bipolar slots span
+  `[-amount, +amount]`; unipolar map sin LFOs to `[0, amount]`. Atomic
+  swap pattern + dealloc queue mirrors `AutomationSnapshot`.
+- **MaskingAnalyzer (offline frequency-masking)** — `ToooT_Plugins/
+  MaskingAnalyzer.swift` is an `actor` that consumes per-channel frozen
+  PCM and produces ranked `[MaskingSuggestion]`. Hann-windowed FFT via
+  `vDSP.FFT`, 10-band log-spaced binning matching `LinearPhaseEQ`,
+  pair-wise mean-energy comparison with configurable threshold + cap.
+  Confidence score 0…1. Pure offline — never touches RT thread.
+- **StemSeparator scaffold** — `ToooT_Plugins/StemSeparator.swift` is the
+  CoreML / Demucs integration scaffold. Streaming AVAudioFile I/O,
+  7.8 s window with 1 s overlap-add crossfade, per-stem WAV writers,
+  cooperative cancellation, MainActor progress callback. Model is loaded
+  from the bundle (`Demucs_htdemucs.mlpackage`); init throws
+  `Error.modelMissing` with a clear searched-path if absent.
+- **MIDIGenerator scaffold** — `ToooT_Plugins/MIDIGenerator.swift` is the
+  same shape: actor, model probe, autoregressive token sampler (softmax
+  with temperature), Magenta "performance" vocabulary decoder that emits
+  `[ClipMIDIEvent]` ready for a `.midi` arrangement clip.
+- **Take-comping UI** — `ToooT_UI/CompingView.swift` renders stacked take
+  rows with painted comp regions; horizontal drag paints a region
+  active for the dragged span on whichever take row was clicked.
+- **AUv3 native-UI embedding** — `ToooT_UI/AUv3PluginViewWindow.swift`
+  ships `AUv3ViewCache` (caches `NSViewController` per plugin instance to
+  skip XPC handshake on reopen), `AUv3HostedView`
+  (`NSViewControllerRepresentable`), `AUv3PluginWindow` (sheet host with
+  generic-parameter-list fallback when the plugin doesn't ship a custom
+  UI). Uses the documented `requestViewController` path so v3 plugins
+  remain out-of-process.
+- **CollabSession foundation** — `ToooT_IO/CollabSession.swift` is the
+  Network.framework P2P transport: `NWBrowser` Bonjour discovery on
+  `_tooot._tcp.local.`, `NWListener` for structural sync (length-prefixed
+  blob channel — Automerge applies once dep is vendored), UDP multicast
+  for live `ParamPacket` parameter drags. Fixed-size 36-byte packet wire
+  format encoded with one `withUnsafeBytes`, no per-call allocations.
+- **DAW-Project XML exporter** — `ToooT_IO/DAWProjectExporter.swift`
+  serialises `Arrangement` (with tracks, pattern/audio/MIDI clips,
+  fades, gains, MIDI notes) to the Bitwig-led MIT-licensed open
+  interchange format. Replaces AAF / OMF as the open path for handing
+  sessions to other DAWs.
+- **Audio + MIDI clip render-path consumption** — `Arrangement` now drives
+  audio for every `ClipKind`, not just `.pattern`. Audio clips trigger a
+  one-shot voice into the unified bank region at clip onset and silence the
+  channel at clip end (`Arrangement.audioClipsStarting/Ending(forChannel:atBeat:)`).
+  MIDI clips carry an embedded `[ClipMIDIEvent]` array (relative beat
+  offsets, `noteOn`/`noteOff`, MIDI note + velocity + instrument) which the
+  render path dispatches at row resolution via the existing voice trigger
+  machinery (`Arrangement.midiEvents(forChannel:atBeat:)`). Pattern-clip
+  override semantics unchanged. UAT 62 verifies onset, end, noteOn, noteOff
+  for both kinds end-to-end.
+- **`RTVoiceScheduler` — opt-in worker pool for realtime voice processing**
+  — spawns N user-interactive `Thread`s coordinated via per-worker
+  `DispatchSemaphore`s plus a single done-semaphore. When
+  `EngineSharedState.realtimeMulticoreEnabled = 1`, the realtime render
+  block partitions active voices across the pool, each writing into its
+  thread-local scratch slot; voice processing + PDC + sidechain peak run
+  in parallel, then the audio thread does a serial vDSP_vsma reduce into
+  the master + bus accumulators. Defaults to off; flip the flag to opt in.
+  UAT 63 covers the scheduler primitive (full dispatch coverage, slot
+  ownership, prompt teardown, post-stop fallback).
+
+### Changed
+- **UATRunner — section reorder + timeouts.** `§15` (60-second openmpt A/B
+  fidelity test) now runs before the summary so its failures count toward
+  the pass/fail tally, and the missing-reference-WAV path falls through
+  instead of `exit(0)`. `§26` (AUv3 plugin discovery + hosting) runs last,
+  bounded by a 15 s discovery timeout and a 30 s hosting timeout — so a
+  cold `AVAudioUnitComponentManager` cache prints a skip-with-warning
+  rather than gating the rest of the suite (was the v2.0.1 carry-over
+  TODO).
+
 ## [2.0.1] — 2026-04-26
 
 ### Added