2.84.5

2.84.5

Fixed

• The experimental label on the per-feed pair-synthesis override no longer
  pushes its dropdown out of line with the rest of the cue tuning controls.
  It now sits at the end of the row, after the "Empty = use global" hint.

2.84.4

Fixed

• react-dom is pinned to the same version as react again. A dependency
  bump moved react to 19.2.8 and left react-dom at 19.2.7, and React
  refuses to run when the two differ, which stopped every frontend test file
  from loading. Production builds strip that check, so the shipped UI was not
  affected, but the mismatch is unsupported either way.
• A podcast whose artwork URL serves the wrong content type is no longer
  refetched on every feed refresh. The download failed, left the cached flag
  unset, and so ran again the next cycle, costing a request and a warning
  every few minutes for as long as the publisher's host stayed broken. A
  failed URL is now remembered for six hours. Changing the artwork URL, or
  refreshing artwork by hand, retries at once.
• Container startup reports when it cannot change ownership of entries under
  the data directory, rather than continuing silently (issue #604). A
  container that dropped CAP_CHOWN fails every entry, and without the
  warning the only symptom was the app being unable to write.
• Sponsor guessing no longer treats a capitalized filler word as a brand.
  The word after an ad transition phrase was compared against the skip list
  with its original case, so a transcript reading "brought to you by The
  folks at ..." produced a sponsor named "The", and "by Today Show" produced
  "Today". Split patterns feed this name to the known-sponsor table, so a
  junk brand could be created there. The comparison now folds case and also
  rejects the shared extraction-failure vocabulary the detector and pattern
  creation already use, plus show-credit verbs such as "produced", which
  follow the sponsor read closely enough to be picked up as brands.

2.84.3

Changed

• Cue-driven ad cutting is now labelled experimental in the UI: the global
  "Create ads from cue pairs" toggle, the per-feed pair-synthesis override, and
  the cue-only processing preset. These are the paths that can cut audio from
  audio-cue evidence alone, without an LLM reading the span. Cue snapping and
  silence snapping are deliberately not labelled, since they only move the
  edges of a cut the detector already found.

2.84.2

Fixed

• An unrecognized WHISPER_DEVICE now transcribes on CPU with a warning
  instead of failing every chunk (issue #605). CTranslate2 accepts only cpu
  and cuda, and the value went to it unvalidated while the CUDA availability
  check and the compute-type fallback were both keyed on the exact string
  cuda, so a near miss such as gpu skipped every guard and left gaps in the
  transcript. The system status endpoint now reports the effective device
  rather than the raw setting.
• The container startup ownership scan says so when it cannot read part of the
  data directory, instead of reporting zero unowned files. A container without
  CAP_DAC_OVERRIDE cannot traverse every entry, and a silent zero reads as
  "nothing to migrate".

Changed

• The GPU image now ships PyTorch built against CUDA 12.9 instead of 12.6, so
  it carries kernels for Blackwell cards (sm_120, the RTX 50 series) and for
  sm_100. Still a CUDA 12.x wheel, so the driver floor stays at 525. The trade
  is that CUDA 12.9 drops the sm_50, sm_60, and sm_70 kernels 12.6 carried, so
  Maxwell, Pascal, and Volta cards will now print an unsupported-architecture
  warning from PyTorch. Transcription is unaffected on those cards because
  CTranslate2 does the transcribing and is compiled separately; PyTorch is
  only used here to report VRAM, and a failed reading falls back to default
  chunk sizing. The GPU image grows by about 1.6 GB.

2.84.1

Fixed

• Ad detections are no longer dropped when a local model answers with the
  singular {"ad": [...]} wrapper instead of {"ads": [...]}. The response
  parser already accepted several wrapper spellings but not this one, so a
  window with real detections parsed as zero ads. Seen with Ollama models
  such as qwen2.5, which have no strict schema enforcement to hold them to
  the requested key. Thanks to @combwizard for the report and the fix
  (PR #603).
• Container startup no longer skips the data-directory ownership migration
  when find cannot read part of the volume (issue #604). Under
  set -o pipefail an unreadable entry failed the whole pipeline, and the
  inline || echo 0 fallback appended a second value instead of replacing
  the count, leaving 0\n0 for the numeric comparison to reject with an
  arithmetic syntax error. Installs on network-mounted volumes (NFS with
  root_squash, for example) hit this most often, which is also where the
  ownership migration matters most.

2.84.0

Added

• Per-feed processing mode is now a single writable preset instead of three
  separate controls. processingMode on the feed PATCH endpoint accepts
  standard, keep_content, skip_detection, or passthrough and writes
  the underlying passthroughEnabled, skipAdDetection, and detectionMode
  columns canonically in one request; sending it together with any of those
  three fields is rejected with a 400. The three legacy fields are still
  accepted on their own for existing API callers. Feed settings now offer
  one "Processing mode" select covering all four states, replacing the old
  detection-mode dropdown plus the separate "Skip ad detection" and
  "Pass-through" toggles.
• New cue_only processing mode: cuts come from cue pairs and previously
  learned ad patterns, with no LLM detection call. It requires the feed to have
  at least one enabled ad_break_start template and one enabled
  ad_break_end template; enabling the mode is rejected with a 400
  otherwise, since a boundary-only cue (the ad_break_boundary type) can
  pair across show content without an LLM pass to catch the mistake.
  Fingerprint, text-pattern, and cross-fetch differential detection still
  run; the LLM detection pass, LLM reviewer, pass-2 verification, and LLM
  redetection are all off in this mode (redetection returns a 409).
• Per-feed cueOnlySafety policy for cue_only feeds. hold_new (the
  default) holds a template's synthesized cuts for review until it has 3
  episodes with a paired start/end match; auto_cut cuts immediately but
  still holds any pair scoring below 0.90 confidence. Holds carry the new
  cue_template_unproven or cue_low_confidence reason.
• Cue template drift detection: the template list now shows each template's
  last match and a "quiet" badge, and a cue_only run fires the new
  Cue Template Quiet webhook and email event when a previously-matching
  enabled template records no matches across the feed's last 5
  telemetry-recorded episodes.
• Per-feed skipTranscription toggle, valid only under cue_only mode.
  Skipping transcription stops generated chapters and removes transcript
  search and VTT subtitles for those episodes, but publisher chapters
  (embedded ID3 or linked Podcasting 2.0 JSON) still get remapped onto the
  cut audio. Runs are labeled "(cue-only)" and, when transcription is
  skipped, also "(no transcript)" in the episode's run list.

Fixed

• Cue-only template eligibility is now enforced on template edits and
  deletes, not just on the processingMode PATCH. Disabling, retyping, or
  deleting a cue_only feed's last enabled ad_break_start or
  ad_break_end template returns a 409 instead of silently leaving the feed
  unable to cut anything; a network-scope template mutation checks every
  sibling feed on the network too, not just the template's owner.