Got the original 'warm during speak' design working after all. The
prior blocker — MLX raising 'There is no Stream(gpu, 0) in current
thread' when a background thread did model.generate() — turns out to
be the same bug Apple's mlx-lm hit and fixed in their server
(ml-explore/mlx-lm#1090). The fix is architectural: route ALL
generate() calls through one dedicated thread, and that thread must
load the model itself (a model can't be safely used from a different
thread than the one it was loaded on).

inference_worker.py — new module. InferenceWorker is a daemon thread
that:

  * Creates its own mx.new_thread_local_stream(mx.gpu).
  * Loads the parakeet model in run(), not on main.
  * Pulls tasks (warmup / transcribe) from a queue and dispatches
    them inside .
  * Exposes wait_ready(), submit_warmup(), submit_transcribe(),
    shutdown(). Used by parakey.py for everything inference-related.
  * Takes a  injection
    point so unit tests can mock MLX out without monkey-patching.

parakey.py — refactored:

  * _initialize() no longer loads the model. It blocks on
    worker.wait_ready() and forwards the worker's status callbacks
    to the menu's load_status text.
  * _start_recording() submits a parallel warmup via the worker if
    WarmupGate's try_begin_warmup() says we're cold. The warmup runs
    on the worker thread WHILE the user is speaking, so by release
    time the GPU is hot.
  * _transcribe_and_paste() submits the real transcribe to the
    worker and waits via a callback. The worker's FIFO queue
    naturally serializes warmup → transcribe — no explicit lock.
  * Menu status reads  (the worker is now the source
    of truth for 'is a warmup in flight').
  * _inline_rewarm() removed — superseded by the parallel approach.

bench_idle.py — updated to drive a real InferenceWorker. Shows the
new ceiling:

  * cold first generate:        ~1.3s
  * steady-state (warm):        ~160ms
  * parallel-warmup pattern:    ~150ms post-release  ← essentially
                                                       equal to
                                                       steady-state

tests/test_inference_worker.py — 6 unit tests with fakes for MLX,
covering:

  * Model loads on the worker thread (not main).
  * wait_ready() blocks until load + startup warmup complete.
  * Load failures are surfaced via .error AND still set _ready.
  * Tasks run FIFO; transcribe waits for in-flight warmup.
  * warming flag is True during warmup and only that.
  * Task errors don't kill the worker.
  * shutdown() drains the queue before exiting.

All 20 tests (14 WarmupGate + 6 InferenceWorker) pass in ~270ms,
and continue to run on Linux CI without MLX installed.