Pipeline: env overrides for endpoints and dirs; voice health check, retries, load-on-demand; refuse speed_factor

_pipeline/settings.yaml

@@ -27,7 +27,7 @@ voice:
   # LOCKED by Ty 2026-10-17 for MVP ("clip B"): Chatterbox, stock voice Michael, native settings only.
   # Never set speed_factor: the server applies it as a time-stretch after generation and ruins the sound.
   api: chatterbox_tts        # chatterbox_tts = native /tts route; anything else = OpenAI-style /v1/audio/speech
-  endpoint: http://voice:8004/tts
+  endpoint: http://voice:8004/tts   # override with env AUDIOBOOK_VOICE_ENDPOINT (CI uses the LAN address)
   model: chatterbox
   voice: Michael.wav
   extra:                     # extra JSON fields sent with every request
@@ -38,12 +38,16 @@ voice:
     split_text: true
     chunk_size: 240          # the server splits long text at sentence ends into chunks of about this size
   timeout_s: 600
+  retries: 3                 # a failed request is retried this many times (5 s, then 15 s, then 45 s apart) before the run fails
+  load_if_needed: true       # if the server is up but its model is unloaded, ask it to load (POST /restart_server) before rendering
+  unload_when_done: false    # true = free the GPU after a render (POST /api/unload); the next run loads it again
   max_chunk_chars: 100000    # one request per paragraph; the server does the sentence splitting (chunk_size above)
   gap_sentence_s: 0.25       # pause between pieces inside a paragraph
   gap_paragraph_s: 0.7       # pause between paragraphs and after cues and labels
 
 ocr:
   enabled: true              # run ocrmypdf only when a page has no text layer
+  required: true             # pages without a text layer and no working OCR = the run fails (env AUDIOBOOK_REQUIRE_OCR=0 to allow)
   language: eng
   min_chars_per_page: 10     # fewer characters than this = no text layer