feat(capture): dictation, personalities, 0.5.0

Ships the Capture release end to end. Global-hotkey dictation with
synthetic paste into the focused app on macOS and Windows, an on-screen
pill across recording / transcribing / refining, customizable push-to-
talk and toggle chords, and an accessibility-permission prompt scoped to
Settings → Captures with inline re-check feedback.

Voice profiles gain optional personalities that power compose / rewrite /
respond actions via a local Qwen3 LLM — shared with refinement, so there
is one local LLM in the app, not two.

Refinement hardened with deterministic Whisper-loop collapse before the
LLM sees the transcript, per-capture flag snapshots for re-runs, and a
ten-transcript evaluation harness across every bundled refinement size.

Version bump 0.4.5 → 0.5.0.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

app/src/components/ServerSettings/GenerationSettings.tsx

@@ -1,116 +0,0 @@
-import { Card, CardContent, CardDescription, CardHeader, CardTitle } from '@/components/ui/card';
-import { Checkbox } from '@/components/ui/checkbox';
-import { Slider } from '@/components/ui/slider';
-import { useServerStore } from '@/stores/serverStore';
-
-export function GenerationSettings() {
-  const maxChunkChars = useServerStore((state) => state.maxChunkChars);
-  const setMaxChunkChars = useServerStore((state) => state.setMaxChunkChars);
-  const crossfadeMs = useServerStore((state) => state.crossfadeMs);
-  const setCrossfadeMs = useServerStore((state) => state.setCrossfadeMs);
-  const normalizeAudio = useServerStore((state) => state.normalizeAudio);
-  const setNormalizeAudio = useServerStore((state) => state.setNormalizeAudio);
-  const autoplayOnGenerate = useServerStore((state) => state.autoplayOnGenerate);
-  const setAutoplayOnGenerate = useServerStore((state) => state.setAutoplayOnGenerate);
-
-  return (
-    <Card role="region" aria-label="Generation Settings" tabIndex={0}>
-      <CardHeader>
-        <CardTitle>Generation Settings</CardTitle>
-        <CardDescription>
-          Controls for long text generation. These settings apply to all engines.
-        </CardDescription>
-      </CardHeader>
-      <CardContent>
-        <div className="space-y-6">
-          <div className="space-y-3">
-            <div className="flex items-center justify-between">
-              <label htmlFor="maxChunkChars" className="text-sm font-medium leading-none">
-                Auto-chunking limit
-              </label>
-              <span className="text-sm tabular-nums text-muted-foreground">
-                {maxChunkChars} chars
-              </span>
-            </div>
-            <Slider
-              id="maxChunkChars"
-              value={[maxChunkChars]}
-              onValueChange={([value]) => setMaxChunkChars(value)}
-              min={100}
-              max={5000}
-              step={50}
-              aria-label="Auto-chunking character limit"
-            />
-            <p className="text-sm text-muted-foreground">
-              Long text is split into chunks at sentence boundaries before generating. Lower values
-              can improve quality for long outputs.
-            </p>
-          </div>
-
-          <div className="space-y-3">
-            <div className="flex items-center justify-between">
-              <label htmlFor="crossfadeMs" className="text-sm font-medium leading-none">
-                Chunk crossfade
-              </label>
-              <span className="text-sm tabular-nums text-muted-foreground">
-                {crossfadeMs === 0 ? 'Cut' : `${crossfadeMs}ms`}
-              </span>
-            </div>
-            <Slider
-              id="crossfadeMs"
-              value={[crossfadeMs]}
-              onValueChange={([value]) => setCrossfadeMs(value)}
-              min={0}
-              max={200}
-              step={10}
-              aria-label="Chunk crossfade duration"
-            />
-            <p className="text-sm text-muted-foreground">
-              Blends audio between chunks to smooth transitions. Set to 0 for a hard cut.
-            </p>
-          </div>
-
-          <div className="flex items-start gap-3">
-            <Checkbox
-              id="normalizeAudio"
-              checked={normalizeAudio}
-              onCheckedChange={setNormalizeAudio}
-              className="mt-[6px]"
-            />
-            <div className="space-y-1">
-              <label
-                htmlFor="normalizeAudio"
-                className="text-sm font-medium leading-none cursor-pointer"
-              >
-                Normalize audio
-              </label>
-              <p className="text-sm text-muted-foreground">
-                Adjusts output volume to a consistent level across generations.
-              </p>
-            </div>
-          </div>
-
-          <div className="flex items-start gap-3">
-            <Checkbox
-              id="autoplayOnGenerate"
-              checked={autoplayOnGenerate}
-              onCheckedChange={setAutoplayOnGenerate}
-              className="mt-[6px]"
-            />
-            <div className="space-y-1">
-              <label
-                htmlFor="autoplayOnGenerate"
-                className="text-sm font-medium leading-none cursor-pointer"
-              >
-                Autoplay on generate
-              </label>
-              <p className="text-sm text-muted-foreground">
-                Automatically play audio when a generation completes.
-              </p>
-            </div>
-          </div>
-        </div>
-      </CardContent>
-    </Card>
-  );
-}

app/src/components/ServerSettings/ModelManagement.tsx

@@ -83,6 +83,12 @@ const MODEL_DESCRIPTIONS: Record<string, string> = {
     'Whisper Large (1.5B parameters). Best accuracy for speech-to-text across multiple languages.',
   'whisper-turbo':
     'Whisper Large v3 Turbo. Pruned for significantly faster inference while maintaining near-large accuracy.',
+  'qwen3-0.6b':
+    'Qwen3 0.6B — smallest of the Qwen3 instruct family. Very fast on CPU, runs at ~400 MB quantized on Apple Silicon. Good for dictation refinement and short completions.',
+  'qwen3-1.7b':
+    'Qwen3 1.7B — balanced size and quality. Handles subtle self-corrections and technical vocabulary better than the 0.6B. Runs at ~1.1 GB quantized on Apple Silicon.',
+  'qwen3-4b':
+    'Qwen3 4B — highest quality local refinement and longer-form reasoning. ~2.5 GB quantized on Apple Silicon, ~8 GB at full precision on PyTorch.',
 };
 
 function formatDownloads(n: number): string {
@@ -411,11 +417,13 @@ export function ModelManagement() {
         m.model_name.startsWith('kokoro'),
     ) ?? [];
   const whisperModels = modelStatus?.models.filter((m) => m.model_name.startsWith('whisper')) ?? [];
+  const llmModels = modelStatus?.models.filter((m) => m.model_name.startsWith('qwen3-')) ?? [];
 
   // Build sections
   const sections: { label: string; models: ModelStatus[] }[] = [
     { label: t('models.sections.voiceGeneration'), models: voiceModels },
     { label: t('models.sections.transcription'), models: whisperModels },
+    { label: t('models.sections.languageModels'), models: llmModels },
   ];
 
   // Get detail modal state for selected model