Add dictation-locale transcription (incl. Russian), rework list capture UI

- TranscriptionSettings/Service route per-locale to SpeechTranscriber or
  DictationTranscriber (union of both supported sets, speech preferred);
  neutral-confidence ranking fallback keeps mixed-engine auto-pick working
- NotesListView: text + voice capture buttons above the list (glassProminent),
  search moved to the iOS 26 toolbar search accessory (.minimize)
- Changelog entries for languages, capture buttons, search placement

Notes/Services/LanguagePicker.swift

@@ -20,6 +20,15 @@ struct LanguageCandidate: Equatable {
 /// side-effect-free is what makes the (undocumented, cross-locale) confidence
 /// calibration unit-testable.
 enum LanguagePicker {
+    /// The confidence assigned to a transcript that carried text but no
+    /// `transcriptionConfidence` attribute at all — the graceful-degradation case
+    /// for `DictationTranscriber`, which need not emit per-run confidence. A
+    /// mid-scale value so such a candidate is ranked by transcript length against
+    /// its peers rather than being zeroed out and always losing to a confidence-
+    /// reporting engine. Two confidence-less candidates then tie here and fall
+    /// through to the length tie-break in `pick`.
+    static let neutralConfidence = 0.5
+
     /// Character-count-weighted mean confidence over an attributed transcript.
     /// Only runs that actually carry a `transcriptionConfidence` attribute
     /// contribute, weighted by their character length; a transcript with no
@@ -37,6 +46,22 @@ enum LanguagePicker {
         return totalCharacters > 0 ? weighted / Double(totalCharacters) : 0
     }
 
+    /// Whether any run in the transcript carries a `transcriptionConfidence`
+    /// attribute — i.e. whether `meanConfidence` measured anything real.
+    static func hasConfidence(of attributed: AttributedString) -> Bool {
+        attributed.runs.contains { $0.transcriptionConfidence != nil }
+    }
+
+    /// The confidence used to *rank* a candidate. Prefers the real weighted mean;
+    /// when the transcript has text but no confidence attribute anywhere (an engine
+    /// that doesn't emit it), falls back to `neutralConfidence` so the candidate
+    /// competes on length instead of being zeroed. An empty transcript scores 0 —
+    /// it can never win against real text.
+    static func rankingConfidence(of attributed: AttributedString) -> Double {
+        if hasConfidence(of: attributed) { return meanConfidence(of: attributed) }
+        return attributed.characters.isEmpty ? 0 : neutralConfidence
+    }
+
     /// The best candidate, or nil when given nothing. Argmax on `meanConfidence`;
     /// a zero-length transcript can only win if every candidate is empty. Ties
     /// break deterministically — higher confidence, then longer transcript, then