mere.run 0.26.0

mere.run 0.26.0

This release contains the complete v0.25.0...v0.26.0 delta: four merged
changes changing 55 files (16,305 insertions and 1,360 deletions). The product
work is centered on making LTX quality and output modality independent,
refreshing the vendored DwarfStar agent runtime, and bringing the public CLI
documentation back into full coverage.

LTX draft/final workflow

#206 separates checkpoint
quality from output modality. video generate now uses --quality draft|final and --output-mode video-only|audio-video, with fast draft
video-only as the default. Final renders use the full dev checkpoint plus the
distilled LoRA without forcing an audio stream.

The former --variant distilled|unified-av selector remains compatible. Bare
and explicit distilled requests now route split MLX installs through the
native V2 standalone-distilled transformer, skip unnecessary audio VAE and
vocoder work, preserve the video-only MP4 contract, and report phase timings.
Canonical full-model inspection also accepts compatible installs that still
carry the legacy video-ltx23-a2vid-mlx manifest ID.

The practical product distinction is now explicit: draft versus final chooses
the speed/quality lane; video-only versus audio-video chooses the deliverable.
Suppressing audio is not treated as a checkpoint-speed optimization.

Refreshed DwarfStar runtime

#204 advances the vendored
DwarfStar runtime used by the DeepSeek V4 Flash premier agent tier from
upstream be434773 to efdadd41. The refresh brings in current Metal kernels,
server request hardening, KV-cache fixes, and session/runtime improvements
while preserving Mere's launch behavior and OpenAI-compatible API contract.

Documentation and CLI accuracy

#203 corrects drift and fills
coverage gaps across configuration, model runtime and benchmarking, API vision
routes, image, speech, text, video, vision, agent workflows, command help, and
site navigation.

Compatibility notes

• Existing --variant distilled|unified-av LTX invocations remain supported.
• Existing compatible full LTX installs with the legacy manifest ID continue
  to inspect and validate, with an explicit compatibility warning.
• DwarfStar's Mere launch and OpenAI-compatible API surface are unchanged.

Release provenance and acceptance

• Exact tag: v0.26.0 at merged main commit
  d5d731dbe40030792989d4d13df624a5d43deca3.
• The hosted tagged-tree gate
  is green across Swift, Linux CLI, compatibility docs, and macOS app-bundle
  checks. The canonical local gate passed 2,065 XCTest cases with 127 expected
  skips and zero failures, plus 13 Swift Testing cases; the docs production
  build also passed.
• macOS: the DMG was built from the exact tag, Developer ID signed, Apple
  notarized, stapled, Gatekeeper accepted, fresh-mounted, and run-smoked. The
  byte-identical DMG is live at
  mere.run/releases/mere-run.dmg.
• Linux ARM64 CUDA: the tar and Debian packages were built natively for
  Blackwell sm_121, checksum-verified, and package-smoked on an NVIDIA GB10
  with CUDA 13.0. The exact extracted launcher completed real Bonsai binary
  and ternary 256x256 image renders through the native quantized_mm CUDA
  backend, reaching 96% GPU utilization in both runs.
• Linux x86_64 CUDA: the tar and Debian packages were built from the exact tag
  on an NVIDIA RTX 3090 (sm_86) with CUDA 12.8 and Ubuntu 24.04. The exact
  extracted launcher completed real Bonsai binary and ternary 256x256 image
  renders through quantized_mm in automatic native CUDA mode: binary in 14
  seconds at 96% peak GPU utilization and ternary in 10 seconds at 100%.

See the complete changelog
and the full tag comparison.