0.2.0

Added

• Natural cloud narration through Unreal Speech, the TTS engine used by Readwise
  Reader. Users can choose the free system voice or supply their own Unreal
  Speech API key and select from its multilingual voice catalog.
• Exact word synchronization for cloud narration from provider-supplied audio
  timestamps, with startup buffering, rolling prefetch, and layered caches.
• Device-local persistent narration caching through IndexedDB, with configurable
  storage limits, LRU eviction, usage statistics, and a clear-cache control.
• Automatic per-note reading checkpoints that resume after reopening a note and
  survive note edits through source offsets, token anchors, and progress fallback.
• Live book statistics for word count, completion percentage, remaining time at
  the current WPM, and estimated 128 kbps narration size.

Changed

• Unreal Speech now groups short sentences into larger chunks, buffers two
  chunks before starting, and keeps a two-chunk lookahead so high-WPM playback
  does not pause at every network boundary.
• Pausing narration now preserves and resumes the active audio session instead
  of synthesizing the same text again.
• Unreal Speech API keys now use Obsidian SecretStorage exclusively. Legacy
  plaintext keys migrate once and are removed from normal plugin data. This
  raises the minimum supported Obsidian version to 1.11.4.
• The session narration cache is byte-bounded at 64 MB on desktop and 32 MB on
  mobile. Persistent cache statistics and LRU eviction now use compact metadata
  without loading or rewriting the stored MP3 collection.
• Unreal Speech can use the full 1000 WPM setting through 5x HTML audio playback,
  while System voice keeps its existing 4x safety limit.
• Periodic checkpoint saves are coalesced and less frequent, while pause, close,
  note switch, restart, and completion still persist or clear immediately.
• Narration failures now show a notice while the visual reader continues on its
  estimated clock.

Fixed

• Clearing narration storage or turning the cache off now clears memory and
  IndexedDB together, and in-flight synthesis cannot repopulate storage afterward.
• Provider text offsets now take precedence over ambiguous spoken-text matching,
  keeping normalized forms such as numbers aligned with the original token.
• Exact-timestamp playback stays frozen during the short seek-restart debounce
  instead of letting the estimated clock skip ahead.
• Changing code-block or frontmatter filtering in an open reader preserves the
  current anchored position rather than resetting to the first word.
• Corrupt or non-finite checkpoint and narration-cache records are rejected before
  they can reach playback or restore logic.