AbstractCore v2.17.0

Recommended defaults now fit the machine they are written on, video joins the model catalog and the
recommendations, and the terminal console reaches parity with the web console's Models and Engines
screens (abstractcore-console 0.3.0).

Added

• Video in the model catalog: Wan2.2 TI2V-5B, T2V-A14B and I2V-A14B (MLX-Gen, Apple silicon), with a
  video tag. Where the run-time memory was measured at the default canvas (resident_bytes), the fit
  verdict uses it instead of the file size (TI2V-5B is a 16.9 GiB download that needs about 58 GiB while
  it generates); the A14B models have no such measurement yet, so their fit uses the file size.
  See Local models.
• A recommended output.video route: MLX-Gen Wan2.2 TI2V-5B (one model for text-to-video and
  image-to-video), written by the first-run seed, abstractcore config apply-recommended (also
  --only video), "download all" and the catalog starter only on Apple silicon where the catalog's
  fit says it fits (about 96 GiB of unified memory or more).
• route_unavailable: {provider, model, reason} on configured capability routes this host cannot run
  (an mlx, mlx-gen or supertonic route on a machine without that engine). abstractcore config defaults prints it on the row. apply-recommended keeps and flags such a route; apply-recommended --force replaces it with this host's recommendation, or removes it (new action cleared, counted in
  the report) where nothing recommended runs. Server and cloud routes are never flagged.
• abstractcore-console 0.3.0 (Rust crate): Hugging Face search (h), set the default text model
  (u), a live downloads view, parallel downloads, a model-type filter (t, video included), engine
  server start/stop (s), continuing a paused install (a, y copies its command), admin-gated
  host-changing verbs for embedding consoles, and route/apply screens that show route_unavailable,
  recommendation_unavailable and the cleared action. See Terminal console.

Changed

• Recommended capability defaults are host-aware for every route, not only text. A route is written
  only where its engine runs: off Apple silicon output.image (MLX-Gen) and output.video stay unset;
  Intel Macs get the Ollama build of the portable text model (LM Studio has no Intel build); hosts
  with neither LM Studio nor Ollama (FreeBSD, 32-bit ARM, RISC-V) get no local text route. Each unset
  route says why (recommendation_unavailable in the grid, action unavailable in the apply plan) and
  names the next step. Apple silicon defaults are unchanged. The per-host table is in
  Centralized config.
• abstractcore engines status checks the CPU architecture: the Ollama and LM Studio install plans
  report available: false with the reason on architectures and operating systems their vendors do
  not build for (see Local engines).

Fixed

• Cloning a voice with a local engine (POST /{omnivoice|qwen3-tts}/v1/voice/clone) returned HTTP 500
  when the voice plugin answered with a plain voice id; it now returns the new voice.

Migrating abstractcore-console from 0.2

0.3.0 is a breaking release for library hosts that embed the console screens (for example the
gateway console): ScreensCtx::new takes a Signal<Access>, ConsoleTransport::start_download
takes expected_bytes, a backend whose job route does not know download groups implements
download_job, and the growing public types are #[non_exhaustive]. The full list is under
"Migrating from 0.2" in console-tui/CHANGELOG.md. The abstractcore-console binary itself needs no
migration.