feat: build from captions.coop/captions-deploy, translation models, relay auto-fill

- images/build.sh builds all three images from one pinned captions-deploy
  commit (submodules pin the captions.coop forks; web/Dockerfile is the UI).
- app: UI image configured by env (landing page, site-config.js), translation
  models downloaded into a volume on start (TRANSLATION_MODELS).
- stt: RELAY_AUTOFILL=source|all|off feeds the capture page's /relay-config.
- Drops the recipe's own nginx.conf and web Dockerfile.

.env.sample

@@ -32,6 +32,15 @@ SECRET_STT_API_KEY_VERSION=v1 # length=48
 ## Metadata-only request logs (no audio/text/tokens)
 #CAPTION_LOG_REQUESTS=0
 
+## Translation models served to browsers (downloaded on start, SHA-256 verified):
+## "all" (~660 MB), a list like "enes,esen", or "none"
+#TRANSLATION_MODELS=all
+
+## Capture-page relay auto-fill: source (default; capture-room tokens only),
+## all (also editor output tokens: service-token holders can then publish to the
+## audience room without review), or off
+#RELAY_AUTOFILL=source
+
 ## Extra exact origins (comma separated, e.g. https://steveseguin.github.io)
 #RELAY_EXTRA_ORIGINS=
 #STT_EXTRA_ORIGINS=