AbstractFramework 0.7.1

A patch release. Pins: AbstractCore 2.20.2, AbstractRuntime 0.8.1 and AbstractGateway 0.8.1; every
other version (the terminal consoles, the browser apps, AbstractAssistant 0.10.0) is unchanged
from 0.7.0.

Fixed

• Email automations never trigger themselves. An automation set to "Email me the result" whose
  filter matched its own result email ran again on every result. Every message the framework sends
  automatically (notifications, sign-in codes, anything an automation sends) now carries
  Auto-Submitted: auto-generated (RFC 3834) and an X-AbstractFramework-Automation header, its
  Message-ID is recorded, and the mail watcher never runs an automation on your own automatic
  mail. When an email arrives also ignores automatic mail from others (auto-replies, vacation
  notices) by default; its auto_submitted: "admit" option turns that off. See the gateway's
  email guide.
• Voice on a light install. Text-to-speech returned an error on a fresh light install of a
  large Mac: the recommended text model there was an MLX build, which the light profile does not
  install, and speech went through the text model. Speech, transcription and image generation no
  longer need a text model, and recommendations only name engines your install has: without MLX
  the Mac's memory tier runs on LM Studio (qwen/qwen3.5-9b below 24 GB, qwen/qwen3.8-27b from
  24 GB). A light install seeded by 0.7.0 keeps its stored route and sees it flagged; set the text
  route in the console (Models, or Settings) to change it.
• The gateway no longer marks a media request's child run failed ("workflow ... not registered
  (after 40 consecutive attempts)").
• A chat's email to your own address runs without asking, and "your own address" is the one My
  email shows (your user email, else the connected mailbox's address).
• A run started through the API without a tool list gets the email tools when you turned on agent
  email tools.
• Moving a model route to another provider drops the old engine's speculation setting (an LM Studio
  route no longer keeps an MLX route's native_mtp).
• A live reply no longer misses its end when the run finished just as the gateway was reading its
  stream.
• install.sh on Linux with an NVIDIA GPU: The same steps by hand now include llama.cpp's CUDA
  build (the listed install line takes its CPU build).

Changed

• The installers do not build the AbstractCore terminal console (abstractcore-console); it is
  listed with each release and installed with cargo install abstractcore-console for a
  standalone abstractcore serve (the gateway's terminal console has its Models and Engines
  screens). See Install.