v3.6.0

[svdC1]

This release hardens the Modal-hosted deploy and adds new host compute modes.
It makes host data durable against a mid-write preemption, adds an offload-less
GPU host and a non-preemptible CPU host, adds a modal logs command, and reworks
the image-version pin into a single deploy-time option applied across every image
command. It also fixes a Windows crash in the modal CLI, several volume-sync
inefficiencies, a stop that could not delete the volume, and a host-mode toggle
that silently did nothing

Breaking Changes

• The image-version pin is renamed. The MIRUMOJI_VERSION config key becomes
  MIRUMOJI_IMAGE_VERSION and the --version command flag becomes
  --image-version, with no compatibility alias, so update any script or
  config that pinned an image version

• The global mirumoji --version (which prints the installed package
  version) is unchanged

Added

• Launcher → An offload-less GPU host. Setting MIRUMOJI_HOST_ON_GPU=1
  (or passing --host-on-gpu) runs the whole host on a GPU with the local
  whisper backend in-process, on the MIRUMOJI_MODAL_GPU type, so there is a
  single always-warm app and no on-demand offload worker

• Launcher → A non-preemptible CPU host. MIRUMOJI_HOST_NONPREEMPTIBLE=1
  (or --nonpreemptible) runs the host on guaranteed capacity at a 3x price, so
  a spot reclaim never restarts it mid-job. It cannot be combined with a GPU host,
  which Modal does not allow

• Launcher → A mirumoji modal logs command that fetches the hosted app's
  recent logs (--tail) or live-follows them (--follow, stopped with
  Ctrl+C)

• Launcher → --image-version is honored by every image command (up,
  pull, render, build, modal deploy), and build now checks out the
  matching v<version> source tag

Changed

• Launcher → The image-version pin is a single deploy-time option resolved
  in one place (flag, then config, then shell, then the installed version) and
  is never injected into a container. See the breaking note above for the
  MIRUMOJI_VERSION / --version rename

Fixed

• Launcher → A Fix-SRT (or any media write) could survive a mid-write
  Modal preemption as a database row whose media file was missing. Media writes
  are now atomic (written to a temporary sibling and renamed into place) and the
  background volume sync is ordered, so the volume never holds a row referencing a
  file that has not been mirrored yet

• Launcher → The modal CLI could crash partway through on Windows when
  its output held a character the legacy cp1252 console could not encode (a box
  border or a Japanese filename), so download-data and log fetching could fail.
  Its subprocesses now encode their output as utf-8

• Launcher → mirumoji modal down --volume could not delete the data
  volume once the app had stopped or failed to stop. A stop failure is now a
  warning, so the volume is still deleted

• Launcher → The host volume sync thrashed the FUSE layer, re-copying a
  transcription's growing scratch audio and the streamed-back converted video many
  times over. Transient scratch is no longer mirrored, and the offload worker's
  result is written atomically, so each file is copied once

---

This discussion was created from the release v3.6.0.