v2.0

[julia-kafarska]

What

Adds a full document / ebook reader to Out Loud, taking it from a "type and speak"
utility to an app that reads whole books aloud. Supports TXT, EPUB, PDF, DOCX, DOC,
MOBI, and HTML, with synchronized word/sentence highlighting, auto-scroll, chapter/page
navigation, and an adaptive sidebar. The reader is an in-place mode: the existing
quick-speak window stays the default, and opening a document grows the window and swaps
to the reader layout. Also ships a Recents library (recently opened documents) and
anonymous, content-free telemetry (PostHog). Tagged as v2.0.

Why

The original TTS path generates the entire input at once and holds all of it in RAM —
the worker keeps every chunk and concatenates one giant waveform, and the renderer
caches every decoded AudioBuffer forever. That's fine for a paragraph and fatal for a
book. There was also no cancellation, and streaming speed was silently broken.

Rather than retrofit that path, this PR adds a parallel, bounded-memory pipeline and
leaves quick-speak untouched. Two principles drive the design:

1. The sentence is the atomic unit for TTS, highlight, scroll, and seek. Blocks
   (paragraphs/headings) are just render containers.
2. Main stays stateless about document content. The renderer owns parsing and the parsed
   document (it has a native DOMParser, and jszip/pdfjs run there), so main is only file
   dialog + byte reads + worker bridge + window resize + recents — no parser dependencies

New dependencies: jszip + pdfjs-dist (renderer), word-extractor (main, for legacy .doc).
Architecture and data model are documented in docs/app/reader.md.

Checklist

• npm run check passes (lint + format + knip + typecheck)
• npm test passes
• I've tested this locally (npm run electron:dev)