fix: made the macos app smaller

lib/bds/embeddings/backends/neural.ex

@@ -32,6 +32,11 @@ defmodule BDS.Embeddings.Backends.Neural do
     * Everywhere else — and as a fallback when EMLX is unavailable or explicitly
       disabled — it runs on optimised native CPU via XLA (`compiler: EXLA`).
 
+  Packaged releases contain only the backend that can run on their platform
+  (`runtime:` conditional deps in mix.exs): macOS releases ship EMLX and no
+  EXLA, Linux/Windows ship EXLA and no EMLX (MLX is Apple-only). The selected
+  backend is therefore started on demand here, never assumed at boot.
+
   The accelerator can be pinned with `config :bds, :embeddings, accelerator:`
   to `:auto` (default), `:emlx`, or `:exla`.
   """
@@ -171,7 +176,20 @@ defmodule BDS.Embeddings.Backends.Neural do
   @doc false
   @spec current_accelerator() :: :emlx | :exla
   def current_accelerator do
-    select_accelerator(configured_accelerator(), emlx_available?(), apple_silicon?())
+    case select_accelerator(
+           configured_accelerator(),
+           emlx_available?(),
+           exla_available?(),
+           apple_silicon?()
+         ) do
+      :none ->
+        raise RuntimeError,
+              "no embeddings accelerator available: neither the EMLX (Apple GPU) nor the " <>
+                "EXLA (CPU) application could be started on this machine"
+
+      accelerator ->
+        accelerator
+    end
   end
 
   @doc """
@@ -179,22 +197,43 @@ defmodule BDS.Embeddings.Backends.Neural do
 
   Prefer the Apple GPU (EMLX) under `:auto` only when it is both available and
   running on Apple Silicon; honour an explicit `:emlx`/`:exla` request, but
-  degrade a forced `:emlx` to EXLA when EMLX is not loaded so a misconfigured
-  host still gets working CPU inference instead of crashing.
+  degrade to the other backend when the requested one is not available so a
+  misconfigured host still gets working inference instead of crashing. Only
+  one backend ships per platform (macOS releases exclude EXLA, Linux/Windows
+  exclude EMLX), so the unavailable side is the norm in production, not an
+  edge case.
   """
-  @spec select_accelerator(:auto | :emlx | :exla, boolean(), boolean()) :: :emlx | :exla
-  def select_accelerator(:exla, _emlx_available?, _apple_silicon?), do: :exla
-  def select_accelerator(:emlx, true, _apple_silicon?), do: :emlx
-  def select_accelerator(:emlx, false, _apple_silicon?), do: :exla
-  def select_accelerator(:auto, true, true), do: :emlx
-  def select_accelerator(:auto, _emlx_available?, _apple_silicon?), do: :exla
+  @spec select_accelerator(:auto | :emlx | :exla, boolean(), boolean(), boolean()) ::
+          :emlx | :exla | :none
+  def select_accelerator(_configured, false, false, _apple_silicon?), do: :none
+  def select_accelerator(:exla, _emlx?, true, _apple_silicon?), do: :exla
+  def select_accelerator(:exla, true, false, _apple_silicon?), do: :emlx
+  def select_accelerator(:emlx, true, _exla?, _apple_silicon?), do: :emlx
+  def select_accelerator(:emlx, false, true, _apple_silicon?), do: :exla
+  def select_accelerator(:auto, true, _exla?, true), do: :emlx
+  def select_accelerator(:auto, _emlx?, true, _apple_silicon?), do: :exla
+  def select_accelerator(:auto, true, false, _apple_silicon?), do: :emlx
 
   defp configured_accelerator do
     config() |> Keyword.get(:accelerator, @default_accelerator)
   end
 
+  # The backend apps are runtime: false deps on the platforms that can't run
+  # them, so releases exclude one of them entirely and neither is auto-started
+  # in dev/test. Availability therefore means "the app actually starts here":
+  # ensure_all_started errors when the app is absent (or its NIF can't load).
+  # Starting the app is also exactly what the chosen backend needs before
+  # building the serving.
   defp emlx_available? do
-    Code.ensure_loaded?(EMLX) and Code.ensure_loaded?(EMLX.Backend)
+    started?(:emlx) and Code.ensure_loaded?(EMLX.Backend)
+  end
+
+  defp exla_available? do
+    started?(:exla)
+  end
+
+  defp started?(app) do
+    match?({:ok, _apps}, Application.ensure_all_started(app))
   end
 
   defp apple_silicon? do