Changelog

AI services (ONNX, LLamaSharp, Whisper.net) allocate model weights through

their own native libraries, bypassing both the managed heap and the Rust

allocator. This made their memory invisible to SubsystemTracker — the

Dashboard showed "—" for Intent Engine, Local LLM, and Speech-to-Text, and

only managed-side allocations for Embeddings/RAG.

SubsystemTracker: add ReportNativeBytes(subsystemId, bytes) and static

convenience method. Allows C# services to report native memory directly,

complementing the existing Rust FFI path (MergeNativeCounters).

EmbeddingService: report ONNX model file size as native memory on load,

clear on unload/dispose.

LocalLlamaProvider: report GGUF model file size as native memory on load,

clear on unload/dispose and when swapping models in EnsureModelLoadedAsync.

WhisperService: report Whisper GGML model file size as native memory on

load, clear on re-init/dispose.

File size is used as the estimate since these libraries memory-map or load

the full model file into native memory. The actual resident footprint may

differ slightly (runtime buffers, KV cache) but file size provides the

right order of magnitude and tracks load/unload transitions accurately.

Desktop ViewModelBase was a 7-line stub with no disposal support, causing

accumulated event handlers, leaked plugin VMs, and inflated subsystem metrics.

ViewModelBase: Mirror SDK pattern — add IDisposable + CompositeDisposable with

Dispose(bool) guard. All 23 desktop VMs now gain Disposables for subscription

cleanup. SyncPairingViewModel's Dispose() updated to override base properly.

SubsystemTracker: Fix rendering metric from cumulative to snapshot. The

Interlocked.Add on UI thread allocation delta caused ManagedAllocBytes to grow

monotonically (950 MB after 5 min idle). Switch to Interlocked.Exchange to

track per-tick delta, matching the pattern already used for runtime.gc.

MainWindow event leaks:

• WireBalloonPositioning: store handler reference as field, unsubscribe

previous before subscribing new (was adding anonymous lambda on every

DataContext change)

• SetupSpeechRecording: track _speechWiredVm, unsubscribe old VM's

TranscriptionReady before subscribing new (was double-subscribing from

constructor + DataContextChanged)

• OnClosing: unsubscribe all window event handlers (Opened, PositionChanged,

PropertyChanged, TitleBarSpacer.PointerPressed), balloon/speech handlers,

dispose _chordTimer, unsubscribe OnMainVmPropertyChanged from VM

MainWindowViewModel:

• EvictPluginCache/ReloadPluginAsync: call Dispose() on removed VMs
• OnWorkspaceChanged: dispose all cached VMs before Clear()
• Cleanup: dispose all cached VMs, unsubscribe SettingsVM event handlers
• SettingsVM: replace anonymous lambdas with named methods

(OnSettingsSwitchWorkspaceRequested, OnSettingsLogoutRequested)

Add a 48x48 circular floating action button (FAB) in the lower-right

corner of the content area panel. Uses the same star glyph as the

sidebar AI icon at 28x28, white on ThemePrimaryBrush background

with drop shadow. Includes gold dot indicator for unseen insights.

Relocate the AI speech balloon from bottom-left (over sidebar star)

to bottom-right, originating above the FAB. Arrow tail flipped to

right-aligned. PositionBalloonOverStar renamed to PositionBalloonOverFab

with right-margin calculation relative to the FAB center.

Sidebar AI star icon remains as a secondary access point.