• Output and context ceilings are now 131072 and 262144 tokens; response and session
  storage grow dynamically to fit. Leaving output length unset (API) or on its new "Auto"
  default (native and web) fills the remaining context instead of stopping at 512 tokens.
• Generated images and audio now play, preview and download directly in the web chat,
  matching what the native app already did; the model downloader's fit estimate now reads
  the console's actual reported memory instead of a fixed figure.
• Fixed the web Settings page rejecting a 256K context selection, several mobile layout
  issues (the Models search/filters row, the chat header, and the conversation history
  drawer), and conversation drafts left without a message no longer clutter history.
• Raised the HTTP server's concurrent-connection limit so an ordinary page load no longer
  risks "Too many active connections."