2.86.2

2.86.2

Fixed

• Circuit-breaker-masked auth outages now freeze the retry budget like
  direct auth failures.
• Title blacklist helper text explains whole-title glob matching and uses
  a neutral example.

2.86.1

Added

• Global "process new episodes first" setting controlling the automatic
  fresh-episode queue boost.

Fixed

• Global show-segments toggle saves immediately instead of waiting for
  Save Changes.
• Per-prompt reset buttons stay visible at default, disabled instead of
  hidden.
• Feed show-segments control is an explicit Inherit/On/Off choice that
  shows the effective value when inheriting.

2.86.0

Added

• Two-tier pattern trust: user-created and community patterns are marked
  "defined"; auto-learned patterns are not. Defined status now drives cut,
  merge, and pass-1 hint behavior throughout the detection pipeline.
• Feed queue priority (#625): normal, high, or low, with automatic boosts
  for episodes published in the last 48 hours and for episodes a user
  explicitly reprocesses. Boosts stack on top of the feed's base priority.
• Per-feed episode title blacklist: case-insensitive glob patterns skip
  queuing and just-in-time processing for matching episodes. A per-feed
  choice serves a skipped episode unmodified in the RSS feed or hides it
  entirely.
• Per-prompt reset buttons in settings (#626): each AI prompt resets to its
  default individually instead of only all-or-nothing.
• A dedicated global Segment actions settings card, plus a global default
  for show-segments (intro/outro/recap) detection that feeds inherit
  unless they set their own value.
• Partial detection: when the AI detection pass fails but pattern and
  cross-fetch evidence already produced cuts, the episode publishes with a
  warning banner and a re-run button instead of failing outright. Window
  counts are exposed in the API, and one automatic low-priority re-detect
  is queued.

Changed

• A user-created or community ad pattern now always cuts its matched
  segment, overriding segment-action keep settings. This applies to both
  the keep partition and the late keep safety net; auto-learned patterns
  are unaffected and still respect keep settings.
• Text-pattern merging: when a defined pattern overlaps an auto-learned
  one, the defined pattern wins ownership of the merged span. The absorbed
  pattern keeps credit for its own matches.
• Pattern learning dedupe: learning a near-identical text now updates the
  existing pattern instead of inserting a duplicate. Only spans that pass
  the learning validation guards count toward match-credit stats.
• Pass-1 sponsor hint amplified: defined patterns now contribute category
  and an opening snippet to the hint; auto-learned patterns contribute
  names only. The detailed list is capped, and neither tier contributes
  match spans or timestamps.
• Default detection prompt: a paid read for another show is now classified
  as sponsor rather than cross_promo. Instances running a customized
  detection prompt need to reset it to pick up this change.
• Differential holds now corroborate against all merged member stages
  instead of a single stage; merged marker labels follow whichever member
  covers the marker.
• Duration-estimated pattern spans are now advisory only: they no longer
  contribute to hold corroboration or label reach. The estimated flag
  survives match merges conservatively, so a merge with any estimated
  member stays advisory.
• Validator merge is now action- and cut-status-aware. A recut can no
  longer restore (un-cut) an ad segment that was already cut.

Fixed

• An aborted verification pass, one where transcription failed or produced
  no segments, no longer reports a clean scan.
• #629: widened matching for JSON-mode rejection responses, plus runtime
  self-correction that retries without JSON mode. A speculative JSON-mode
  fallback for unprobed endpoints now persists only after it succeeds,
  reverting on any retry failure.
• #631: LLM responses wrapped in a segments field now parse ads that are
  missing a type field.
• Merge propagation: a defined pattern folding into a claude-first marker,
  or an estimated span promoted to a merged marker's stage, now carries its
  flag through detector and validator merges instead of losing it mid-fold.
• Auth-class LLM failures (invalid key, forbidden, unauthorized) no longer
  consume the episode's retry budget.
• Markers persisted by a failed processing run no longer display as cut
  ads.
• TTL caches are now bounded with a size cap and eviction, rather than
  growing without limit (#621).