2.76.1

2.76.1

Changed

• The update panel in Settings shows the release channel next to the
  running version ("Running 2.76.1 (edge)") and tucks the channel
  selector, daily-check toggle, check button, and changelog link behind
  an "Update settings" disclosure, collapsed by default.

2.76.0

Added

• Six ad-detection tunables in Settings > Ad Detection: a verification-pass
  miss hold floor (default 0.60) and autocut threshold (off by default),
  pattern-learning confidence floors for short and long ads (default 0.85 /
  0.92), and a differential correlation ceiling (default 0.60) plus a
  differential hold minimum length (default 10s, 0 disables it). See
  Configuration > Detection Tuning.
• Standalone pass-2 verification misses (ads pass 2 finds that overlap no
  pass-1 marker) are no longer silently discarded. They now hold for review
  with hold reason verification_miss (a "Verification catch" chip in Held
  for Review) once they clear the new hold floor, or cut automatically when
  the opt-in autocut floor is enabled and cleared. Below both floors the
  miss is still dropped, but the drop is now logged with the sponsor,
  confidence, and floor it fell short of.
• Cue fusion joins the cross-fetch differential stage on feeds with cue
  templates configured. A matched template cue on either edge of an
  otherwise-uncorroborated differential candidate now corroborates it (cuts
  instead of holding), a candidate bracketed by a break-start/break-end cue
  pair corroborates the same way, and the refetch is scanned for the same
  cues as the primary download so a shared cue re-anchors the comparison
  timeline against fetch-to-fetch drift. Cross-fetch differential detection
  is now significantly more accurate on cue-templated feeds; see Audio Cue
  Detection and How It Works > Cross-Fetch
  Differential.
• Episode Processed webhooks and emails now carry ads_held and
  ads_not_cut counts (episode.ads_held / episode.ads_not_cut template
  fields); the email adds an "Ads held for review" and/or "Detections not
  cut" row only when either is nonzero.

Changed

• The cross-fetch differential stage now measures rather than assumes:
  every silence-delimited block in the run file is probed against the
  refetch and carries its own measured correlation, replacing the previous
  hard-coded 0.0 for unprobed gaps. A region only becomes a differential
  candidate when its measured correlation is at or below the new
  correlation-ceiling setting, and a candidate close to the boundary gets
  one retry with a widened search window before being judged. Old stored
  differentials (which hard-coded corr 0.0) still qualify, so recuts of
  previously-processed episodes behave the same as before.
• An uncorroborated differential candidate shorter than the new
  hold-minimum-length setting is dropped instead of held for review;
  corroborated candidates are unaffected regardless of length.
• Rejecting a differential detection, held or not, still blocks that same
  episode-region from resurfacing, but no longer creates cross-episode
  false-positive text: previously, rejecting one of these could
  suppress legitimate future matches feed-wide, since the detection was
  never a confirmed false positive from a real detector. A one-time
  startup backfill deactivates existing cross-episode false-positive text
  traceable to a differential rejection (episodes with an ambiguous id
  are skipped rather than guessed at); nothing is deleted. Corrections
  now also carry a source_hold_reason field recording which hold
  reason, if any, produced them.

Fixed

• Sponsor labels no longer pick up LLM reasoning prose or a bare "...
  segment" name instead of the actual advertiser; a sanitizer strips both
  before a label reaches a marker. Two accepted markers describing the same
  ad read that overlap by 80% or more of the shorter span's duration now
  merge into one marker spanning their union, instead of showing up as
  separate, duplicate detections.
• AdDetector.learn_from_detections now initializes its dependencies on a
  cold detector, matching its three sibling methods; previously a cold
  detector (one that had not yet processed anything) silently no-opped
  instead of learning from the detection it was given.
• SMTP send failures now log the full exception traceback instead of just
  the message, for easier troubleshooting from container logs (issue
  #571).
• The startup banner's ASCII art no longer gets dropped by journald-backed
  log drivers (podman): the banner record now starts on its first art line
  instead of a leading blank line, which some drivers treat as an empty
  record and discard (issue #567 discussion).