Release 0.5.0

[dkackman]

Release 0.5.0, cut by run-release.sh from develop @ c693980. Gates on the release issue #515.

Most of this range landed on develop without PRs, so the auto-generated
notes name only the Mac PR (#470) and a deploy-script PR (#476). Paste this
section into the GitHub release body once the tag has published (gh release edit v0.5.0 --notes-file ...).

Breaking and behaviour changes

• templates/ltx2/two-stage no longer has full_width/full_height. They
  never changed the output, which is always exactly 2x width/height. A
  call that still passes them is refused as an unknown variable (ltx2/two-stage: full_width/full_height silently ignored unless exactly 2x base; 480x288 + full 960x544 renders 960x576, no warning #506).
• Every LTX-2.5 template refuses a width or height that isn't a multiple
  of 32 at validate and at run_workflow, before the pipeline loads (ltx2/two-stage: width/height not divisible by 32 passes validate, fails at run after the pipeline load (no constraint declared, unlike num_frames) #505).
• The MiniMax-H3 Ref2VA VRAM ceiling is now projected per step, after
  for_each expansion, and adds 1 GB for each non-null reference. At
  1344x768 on a 24 GB card the limits are 243/209/175/141 frames with 1/2/3/4
  references. Calls that used to pass are now refused, for example 209
  frames with 3 references. A for_each refusal names the member
  (Member 'shot@...'). Eight more Ref2VA templates declare the ceiling
  (Ref2VA VRAM ceiling misses custom workflows and has no per-reference term: 209 frames x 3 refs at 1344x768 projects ~22.5 GB, needed ~29 #479/#479 stage A: per-step, post-expansion H3 VRAM ceiling with a per-reference term #501).
• A workflow with no vram_estimate gets a vram_projection_inherited
  warning, not a refusal, when the catalog template that loads the same
  pipeline would be over its ceiling (#479 stage B: inherit the H3 VRAM ceiling by pipeline identity, as a warning #502).
• The VRAM ceiling is checked against the serving device's own cost entries
  (or its measured capacity), not every CUDA card in the catalog. A 64 GB
  Mac is no longer refused against the RTX 3090 entry (Mac (MPS) support: CUDA templates run unchanged on Apple Silicon #470).
• asset:, prompt: and output: references written directly into a step
  are resolved at validate and at submission. A stored workflow with a
  literal reference to a missing file now fails validation (#485 stage A: attribute_voices task + assessment flag + demucs dependency #494).
• validate_workflow refuses a for_each item's reference that resolves to
  null, which the run already refused (validate_workflow passes a for_each item's null from_file media reference; the run's prepare step rejects it #478). It also refuses a frame-size
  mismatch between the asset:/output:/path inputs of concat_videos and
  dissolve_videos (dissolve_videos/concat_videos: size mismatch between asset: shots passes validate, fails at run, and no task can resize a video to fix it #504).
• A transcribe_audio step with timestamps must set content_type: application/json (transcribe_audio timestamps with a text/plain result validates clean, then fails at save with 'write() argument must be str, not list' #498).
• A weight_name (a LoRA's, an IP adapter's) must be a relative file inside
  the model repo. An absolute path, a backslash, a drive or a ./..
  segment is refused at validate. Media path and glob refusals no longer
  list the server's directories.
• previous_result:<step>.<key> for a key that no result carries is an
  error. It used to run zero iterations and save nothing.
• templates/minimax/video-with-audio-768p runs its turbo LoRA at its
  trained strength. It had been running at 16x, so its output changes (templates/minimax/video-with-audio-768p runs its turbo LoRA at 16x its trained strength (lora_alpha: 128) #468).
• On macOS, parallel checkpoint loading is off by default, and so is SDNQ
  quantized matmul on MPS. Group-offload CUDA streams are dropped, with a
  warning, when the onload device isn't CUDA or XPU. An explicit
  HF_ENABLE_PARALLEL_LOADING still wins (Mac (MPS) support: CUDA templates run unchanged on Apple Silicon #470).
• A for_each whose members produce the same shot name now warns
  shot_name_collision (concat_videos: joining two cuts whose inner shots share names yields duplicate names in media.shots (shot@accuse twice), no warning #508). assess_output has a new shot_dead_air
  finding for silent gaps inside a shot, and concat_videos marks its joins
  as hard cuts, so seam_frame_jump no longer fires on a deliberate cut
  (assess_output doesn't flag in-shot dead air: H3 dialogue shots hold 0.5-2 s at -70..-84 dBFS between lines #465/seam_frame_jump fires on ordinary hard cuts (ratios 25.7-28 vs threshold 25) - 6/6 findings were deliberate framing changes #466).

New

• Apple Silicon (MPS): the CUDA templates, LTX-2.5 and minimax/music
  included, run unchanged. Memory figures and the chip name are real, and
  torch's CPU-fallback warning is shown (Mac (MPS) support: CUDA templates run unchanged on Apple Silicon #470).
• The attribute_voices task names which reference voice sings or speaks
  each line or window, by timbre. It uses demucs separation and ECAPA
  embeddings, and demucs is now a dependency. list_tasks marks the
  read-only probes with an assessment flag (idea: attribute_voices - which reference singer sings each lyric line, by timbre (pitch mislabelled 6 lip-sync shots) #485/#485 stage A: attribute_voices task + assessment flag + demucs dependency #494). The
  minimax-music3 skill points to it for multi-singer songs (#485 stage B: minimax-music3 skill note naming attribute_voices #495).
• normalize_audio(limit=true): a true-peak look-ahead limiter that reaches
  target_lufs past a loud transient. It warns limiter_heavy past 6 dB of
  reduction and target_lufs_capped (limited: true) past 12 dB
  (Limiter for normalize_audio: reach target_lufs past the peak ceiling (split from #467) #474/#474 stage 1: normalize_audio(limit=true), a true-peak look-ahead limiter #496).
• assemble-and-score takes target_lufs (Episodes can't be loudness-matched: assemble-and-score doesn't expose target_lufs; skill should match downward #467) and limit (#474 stage 2: expose limit on assemble-and-score and in the series-episodes skill #497, still
  open). The series-episodes skill says to match episodes downward to one
  series target. Unfinished (#474 stage 2: expose limit on assemble-and-score and in the series-episodes skill #497): with limit: true the mix holds
  -3 dBTP, but the AAC-muxed film can land up to about 1 dB above it.
• transcribe_audio(timestamps="segment"|"word") returns {text, chunks}.
  Unset, it still returns plain text (transcribe_audio discards the timestamps Whisper already returns past 30 s; add opt-in segment/word timestamps #483).
• templates/ltx2/extend-clip takes a clip to extend an existing clip.
  The opening isn't generated when one is given (ltx2/extend-clip can't extend an existing clip: the opening is always generated, and there's no input-clip variable #446).
• save_workflow, delete_workflow, upload_asset, delete_asset and
  list_assets take workspace= for one call. Their replies name the
  workspace they acted on (save_workflow / delete_workflow take no workspace= and act on the shared pin: deleted the wrong workspace's copy #463).
• A LoRA with model_name: null is switched off and warns lora_disabled.
  It used to validate clean and then fail after the model load (Nulling a template's lora variables validated clean, then died on float(None) after the model load; null model_name is now the off switch #469).
• get_guide(section=...) reaches a ### subsection and returns its
  parent_section (get_guide: section= can't reach a ### subsection (for_each), and the refusal doesn't say where it lives #503).
• upload_asset's refusals give a working curl to POST /api/uploads
  and state the 200 MB limit. The guide explains that for_each
  item-level from_previous_result must name the member, as
  slice@<entry> (upload_asset's refusal and the guide never name POST /api/uploads: a remote client's reference images never reached the server #481/Guide: item-level from_previous_result inside for_each must spell slice@<entry>; it does not auto-pair like step level #482).
• A running job's manifest.json is rewritten after each step, so finished
  for_each members show up before the run ends (A running job's manifest.json has steps: [] until the job ends; finished for_each members are only findable via list_gallery #480).
• SECURITY.md: report vulnerabilities privately, not as issues.

Fixes

• templates/ltx2/extend-clip frees the opening's pipeline before the
  extension loads, so it no longer holds two LTX-2.5 stacks on the GPU
  ([templates-ui-packaging] templates/ltx2/extend-clip holds two full LTX-2.5 stacks on the GPU on its default path #523).
• concat_videos/dissolve_videos load a {"location": ...} entry in
  videos (concat_videos: a {"location": "asset:…"} entry in videos validates clean, fails at run ('Cannot extract frames from a dict') after the paid step #510).
• A modular step whose outputs include latents no longer crashes at save
  (run_workflow: result on a modular step with latents output crashes save (append_data requires ndarray) #507). pair_audio unwraps a pipeline's batch of one video.
• transcribe_audio works on clips over 30 s, and a non-Whisper model on a
  long clip still returns text.
• segment works again: the SAM2 KeyError is fixed, and the
  GroundingDINO query is written the way it scores. Dead default input URLs
  in several templates are replaced (Mac (MPS) support: CUDA templates run unchanged on Apple Silicon #470).
• upload_asset(file_path=...) accepts the writable shared asset library
  (SE-F027: upload_asset(file_path=...) refuses the writable common/assets shared library, not just out-of-root paths #448).
• host_memory_job_peak_rss_mb is a running max (C-F047: host_memory_job_peak_rss_mb tracks current rss, not a running max #457).
• dw.serve exits on SIGTERM with an MCP client still connected (dw.serve never exits on SIGTERM while an MCP client stays connected (uvicorn graceful shutdown has no timeout) #477).
• The shot_dead_air finding names the room-tone remedy. The
  normalize_audio warning no longer says that gaining down always succeeds
  (assess_output: shot_dead_air finding's says omits the room-tone bed remedy (#465 item 4) #491/assemble-and-score / series-episodes: 'gain down always succeeds' is false when the mix already peaks above -3 dBFS; target_lufs_capped wording unclear for that case #492). The docs describe dead_air_floor_dbfs correctly ([mcp-and-docs] TASKS.md says shots[].dead_air_floor_dbfs is the DEAD_AIR_FLOOR_DBFS threshold; it is the quietest window's level in the run #519).
• The minimax-h3 skill carries the 24-shot field report's prompting rules
  (Skill notes from a 24-shot H3 musical: props vs actions, landmark direction, two-person lip-sync target, crowd age, reference leak, native 1344x768; Music3 ignores per-section singers #484).