2.83.1

2.83.1

Added

• Per-feed "Skip verification pass" toggle (#599). The pass-2 verification
  scan re-transcribes the cut audio and runs a second full detection sweep
  over it, which roughly doubles the ad-detection LLM spend on every episode.
  It had no toggle at any scope. Feed settings now has a per-feed switch that
  skips it, so a feed whose first pass is already reliable pays for one
  detection sweep instead of two. Held differential detections that pass 2
  would have confirmed wait for a human instead. The toggle reuses a column
  of the same name from the old two-pass era, so upgrading resets any value
  left over from it and the toggle starts off on every feed. A skipped run
  records no verification result rather than a zero, which would have read as
  a clean second scan, and the run list labels it "(no verification)".

Fixed

• An oversized learned-pattern span could swallow a precise LLM detection
  and then hold the whole thing, leaving the ad in the audio. A pattern
  span is minted from a stored average duration; on one episode it ran 62
  seconds past the actual read, the exact 0.98-confidence LLM detection
  inside it was dropped as already covered, and the bloated marker was held
  because its edges sat far from any splice evidence. When exactly one
  high-confidence cut-resolving LLM detection sits inside a pattern span
  that materially overshoots it, the span now adopts the LLM bounds.
• A feed body cut mid-character parsed as a short but valid feed. The
  truncation detector matched the parser's missing-element errors but not
  "not well-formed (invalid token)", which is what a partial multibyte
  sequence produces; it is now treated as truncation, so the stored episode
  list survives.
• HTTP 404s from unknown-path scanner probes logged at warning in the
  access log even after the application-level line was demoted; both now
  log at info.
• The waveform editor opened on the wrong part of the episode at maximum
  zoom when reached from the Patterns page. Neither the Ad Review nor the
  Detected Ads tab passed the episode length, so the editor fell back to
  assuming a six-minute episode and clamped the window to the end of that
  imaginary range. The detections API now carries the episode duration and
  both tabs pass it through.

2.83.0

Added

• Cover-art badge corner is configurable (#600). The MinusPod badge was
  pinned to the bottom-right, where it covered the network logo on some
  shows. A new "Badge position" select under Settings > Cover Art picks
  bottom-right (default), bottom-left, top-right, or top-left, and
  ARTWORK_BADGE_POSITION seeds it on a fresh deploy. The corner is part
  of the badge cache key, and changing it clears the feed validators so the
  served feeds re-render with a fresh cover URL instead of waiting for the
  publisher to change something.

2.82.2

Added

• Per-feed switch to serve MinusPod episode ids as RSS item GUIDs (#598).
  The served feed previously mixed GUID schemes: upstream entries kept the
  publisher's GUID while DB-appended episodes used MinusPod ids. New feeds
  serve MinusPod ids from the first fetch; existing feeds keep upstream
  GUIDs unless the new "Serve MinusPod episode IDs" toggle in feed settings
  is turned on. Caveat: switching an already-subscribed feed changes every
  item GUID, so podcast apps treat each episode as new one time.

Fixed

• Settings card carets behaved differently per card: half rotated in place
  and half jumped down a line on expand, because an appearing subtitle grew
  the header under a centered caret (#597). The caret is pinned to the
  title line, so every card rotates in place.

2.82.1

Added

• Editable chapter generation prompt: the topic-detection prompt is now a
  settings-backed template (chapterPrompt) with placeholders for the split
  count, segment range, continuation/description/hint blocks, and transcript,
  plus a chapterPromptOverride appended at run time. Editable in Settings
  under Prompts and included in the prompts reset; left at the default it
  renders the same prompt as before.

Changed

• The four Chapter Density tunables (target chapter length, transcript window,
  maximum chapters, shortest chapter) moved from LLM Tunables to the
  Transcripts & Chapters card, next to the Generate Chapters toggle, with
  their own Save button. The settings API payload is unchanged.

Fixed

• Feed descriptions were stored clipped to 500 characters, which read as
  visible truncation on the feed page once the UI stopped line-clamping
  them (#596). The stored bound is now 10000 characters, which no real
  description reaches; the full text appears after the feed's next refresh.
• A section header action, such as the model-list refresh, no longer
  renders a button nested inside the section's toggle button.

2.82.0

Added

• Detected Ads tab on the patterns page: browse every ad that was cut,
  across all feeds, filtered by podcast, segment category, and search,
  sorted by date, confidence, or podcast. The header leads with total time
  cut, then detection count, distinct sponsors, and distinct podcasts, with a
  per-category breakdown. Rows play the span, open it in the waveform editor,
  reject it (which recuts from the retained original so the audio comes back),
  or split it.
• Segment category filter on the patterns list and the ad review tab, with a
  separate Uncategorized option for markers no detection stage classified.
• Split a merged multi-sponsor ad marker into N single-sponsor ads
  (issue #563), from a Detected Ads row or from inside the boundary review
  editor. A new split-candidates endpoint finds ad-transition phrases in
  the marker's transcript and maps each to a timestamp, so the editor opens
  with dividers where a split would cut. Dividers are draggable and
  keyboard-operable, each resulting ad takes its own sponsor, and saving
  replaces the one marker with N markers and mints one pattern per piece.
  Every piece must clear the minimum ad duration or the split is refused and
  the marker is left untouched. This closes the issue: the automatic
  multi-sponsor split shipped in 2.70.0 and the editor affordance for held and
  not-cut rows in 2.72.0.
• Configurable chapter density via four new LLM tunables: target chapter
  length, transcript window per detection call, maximum chapters, and shortest
  chapter. Available in Settings under LLM Tunables and via the settings API.

Fixed

• Episode discovery silently dropped episodes when SQLite was busy. The upsert
  ran on a bare connection, so the deferred transaction upgraded to a write
  lock at the first insert, which fails immediately with "database is locked"
  rather than waiting on busy_timeout, and a per-episode handler swallowed it.
  A contended refresh logged one warning per episode and reported success
  having discovered nothing. The batch now holds an immediate transaction,
  lock failures abort it so the caller retries the feed, and per-row data
  faults collapse into one summary warning.
• A truncated feed response was parsed as a short but valid feed, so episodes
  vanished from that feed on every refresh while the circuit breaker recorded a
  success. The shared streaming reader enforced only an upper size bound and
  returned whatever arrived; it now rejects a body that ends short of its
  declared Content-Length, and both RSS fetch paths treat that as a fetch
  failure so the stored episode list survives. A feed document whose parse
  error signals truncation is likewise rejected rather than parsed partially.
• Chapter generation capped every episode at 6 topic boundaries regardless of
  length, and a single detection call over a long transcript clustered the
  boundaries it did return, leaving hour-long chapters at the end. Detection
  now runs per transcript window with the previous window's title carried
  forward, and a failed window no longer discards the others' results. The
  boundary parser also accepts timestamps past 99:59, which windowed
  detection produces on every episode longer than that, and show-notes
  timestamp anchors now reach every window instead of only the first.
• Transcription repeated the same CUDA out-of-memory failures on every
  episode. Batch size came from audio duration alone, so on a card where the
  top tier never fits, each episode paid two OOM failures before settling on a
  size that worked and then discarded that result. The working size is now
  remembered per device.
• The scheduled-backup warning about provider secrets fired on the wrong
  condition. It warned when a master passphrase was configured, which is
  exactly when those secrets are already encrypted at rest, and said nothing
  in the case where they are stored in plaintext. It now warns when plaintext
  secrets are present, names the remedy, and logs once per process rather than
  once per backup.
• Podping node failures logged a warning on every attempt, so a node that
  stayed down printed once a minute indefinitely even though failover was
  working. A node now warns on its first failure, repeats drop to debug, and
  losing every node escalates to an error, which is the case that loses
  pings.
• Unauthenticated requests for feed assets and probes for unknown feed slugs
  logged at warning level, putting routine internet background noise at the
  same level as problems worth acting on. Both drop to info. Failed login
  attempts stay at warning.
• Waveform editor boundary pins declared themselves as sliders to assistive
  technology but had no keyboard handling and no minimum or maximum, so they
  were announced as operable and were not. Arrow keys now nudge them, Shift
  takes a coarser step, and both bounds are exposed.
• Scope, Origin, and Source filter labels on the patterns list were not
  associated with their controls, so none were reachable by label.
• Boundary heuristics could undo a cue-anchored edge and swallow neighbouring
  audio, leaving a real ad in the episode. A text-pattern match landed on an
  ad and cue snapping anchored its end to the feed's learned ad-break
  stinger, then phrase refinement and the content-extension walk moved both
  edges into surrounding speech, across a segment the detector had decided
  to keep. The bloated marker was held for review because its edges were no
  longer near any splice evidence, and a precise LLM detection of the same
  ad had been dropped as redundant, so nothing was cut. Cue-snapped edges
  are now pinned against phrase and content heuristics, the extension walk
  stops at a neighbouring detection's span including segments the feed
  keeps, and a learned template cue firing at a marker edge now counts as
  splice evidence for the veto.
• A marker could end past the end of the file. The validator clamps bounds
  to the episode duration, but protected merge bookkeeping recorded before
  the clamp let the reviewer re-expand the edge past it; the protected
  bounds are now clamped the same way.
• Requesting cover art for a feed slug that does not exist created that
  slug's directory on disk, so scanner probes left orphan directories for
  the refresh cycle to clean up. Artwork reads no longer create anything.
• The corrections endpoint documentation described a request shape the
  server never accepted; it now documents the real one, including the
  trimmed-approval and split fields.