feat(books): add read-aloud generation

Generate a synced read-aloud EPUB from a book's EPUB and audiobook through an
optional STAlign worker (Docker on CPU, or native with MLX on Apple Silicon),
with review, re-sync, replace in place, restart recovery and live progress on
the Files tab. Book details also stack the ebook and audiobook covers when
both exist. See docs/READ_ALOUD.md.

.env.example

@@ -47,6 +47,23 @@ APP_URL=https://bookorbit.example.com
 # Node.js JavaScript heap limit in MB. Raise this for very large libraries, 250K+ books.
 NODE_MAX_OLD_SPACE_SIZE=2048
 
+# Optional read-aloud generation. Enable the compose profile and use the same random token for
+# BookOrbit and its private STAlign worker. The first job per language downloads a speech model
+# (about 190 MB) into data/app/read-aloud. Generate the token with: openssl rand -hex 32
+# COMPOSE_PROFILES=read-aloud
+# READ_ALOUD_WORKER_URL=http://read-aloud-worker:8787
+# READ_ALOUD_WORKER_TOKEN=
+# READ_ALOUD_WORKER_IMAGE=ghcr.io/bookorbit/bookorbit-read-aloud-worker:latest
+# READ_ALOUD_MAX_OUTPUT_BYTES=21474836480
+# Folders the worker can read books from, comma-separated. Add a read-only mount for each extra
+# library folder to the worker service as well.
+# READ_ALOUD_LIBRARY_ROOTS=/books
+# The worker image runs on CPU. A native worker on Apple Silicon uses MLX when this is auto; cuda is
+# only for a native Linux worker with CUDA 12 and cuDNN 9, and is checked at startup.
+# READ_ALOUD_CTC_DEVICE=auto
+# READ_ALOUD_CTC_DTYPE=auto
+# READ_ALOUD_CTC_BATCH_SIZE=
+
 # --- Optional overrides (uncomment to set) ---
 
 # If you already run your own PostgreSQL, set DATABASE_URL and remove the