v0.4.0

Monitoring and scheduled search: the release where the monitored flag starts
doing real work. Add a series, touch nothing, and episodes arrive — searched on
a broadcast-aware cadence, graded against your quality profile, grabbed,
imported, and remembered when they fail. Automation ships off by default;
enable it under Settings → Automation.

Added

• Scheduled search sweep. A wanted-search job periodically searches
  monitored series with outstanding wanted items once their episodes have aired
  (an unknown air date counts as searchable — absence is normal, not an error),
  runs the same match-and-score path as manual search, and grabs only a fully
  eligible release: matched, clear of excludes, over the profile floor. No
  eligible candidate means grab nothing — "nothing yet" is a correct outcome.
  Empty passes back off exponentially from an hour toward a day, clamped so a
  weekly show is still searched at broadcast however many empty passes came
  before; a new episode airing resets the clock, and integration edits in
  Settings apply on the next pass without a restart.
• Recent-feed polling. The sweep costs one indexer query per series, so its
  interval must be long; the Torznab recent feed inverts that — one request
  returns the newest releases across the endpoint, matched against everything
  wanted at once. The feed is the hot path, polled on a short interval, and the
  sweep becomes the safety net. A new release for a wanted, aired episode is
  grabbed on the next poll, not the next sweep. An indexer without a feed
  degrades to sweep-only as a supported configuration, and a quiet poll costs
  one request.
• A pinned group can now mean "wait for," not just "win when present." The
  sweep holds another group's release for a delay window measured from the
  episode's broadcast, so the pinned group's slower release gets its chance
  before automation settles. Global default plus a per-series override; a
  manual grab is never held.
• Automation controls. A global auto-search toggle and the default pin
  delay, persisted and editable live in Settings → Automation — flipping the
  switch takes effect at the next tick, off or on, without a restart. Off
  until you enable it, deliberately: an install that has never configured an
  indexer must not start grabbing on its own. A monitored series page says so
  when automation is globally off, since monitored now means "will be grabbed
  automatically," not just "appears on the calendar."
• Failure memory: a release that fails is not chosen again. The sweep's
  determinism made its own bug — a failed download reverted its item to wanted,
  the next pass re-derived the same ranking, and the same release won forever.
  Failed releases are now blocklisted per series with an escalating expiry
  (24 hours, then 7 days, then permanent): repeat failure of the same release
  is the only signal separating a dead release from a bad day, so first
  failures stay recoverable while a proven-dead release converges to permanent.
  Identity is info hash and normalized title, because Torznab feeds often
  omit the hash. Automation degrades to the next-best eligible release;
  blocked releases appear in the series History tab with the reason, an
  unblock, and bulk clear actions — and a manual grab is never refused and
  never leaves a block behind.
• An environmental-fault breaker guards the failure memory itself. A full
  disk or a client reaping torrents fails a different release every pass —
  the escalation ladder never triggers, and the candidate pool drains into the
  blocklist one entry at a time. Failures across distinct items library-wide
  now trip a breaker that suppresses recording (never grabbing) until the
  window passes; Settings says so in words, and clearing the library blocklist
  also closes the breaker so a fixed fault doesn't wait out its window.
  Refused adds are remembered too, when the client can attribute the failure
  to the release itself — an unreachable client blocklists nothing.
• Background jobs status in Settings. Each runner job's last run, duration,
  and last error (styled as one), refreshed live — "did automation actually
  run, and did it fail?" no longer requires reading server logs.
• Season packs are matched honestly, and automation declines them. Packs
  previously surfaced as unmatched with a reason; they now match their items
  and are declined by automation with an explicit ineligible reason instead —
  the importer can only defer a multi-episode payload, and unattended grabbing
  must not volunteer for a state it cannot finish. A manual grab of a pack
  still succeeds, landing in the deferred flow as before. Per-file batch
  import is tracked for a later release.

Fixed

• Search terms with unicode punctuation found nothing. AniList titles carry
  typography that never appears in release names (×, ☆, ・), so a
  HUNTER×HUNTER-class title searched verbatim died before matching ever ran —
  and unattended search would have inherited the silence. Queries are now
  sanitized (× → x, separators to spaces), zero results fall back to the
  English then native title variants, and matching transliterates instead of
  deleting, so the romaji variant matches real releases on its own.
• Dual-titled scene releases misparsed the alternate title as the release
  group. The trailing (Romaji Title, Multi-Audio, Multi-Subs) parenthetical
  became the group, so group ranking, blocking, and pinning silently never
  applied to a whole common release family. An implausible group is now
  distrusted and the real scene group recovered from the codec-dash pattern.
  The same family's no-episode-title variant also lost its source and codec to
  a misfiled tag run; both now survive.
• Saving any Settings section no longer blanks the Authentication card. The
  section writers returned a settings body without the auth block and the UI
  cached it whole; all writers now share one complete response.

Internal

• Search, decide, and grab moved into one core package (internal/core/ acquire), shared by the manual routes, the sweep, and the feed poll — one
  matcher, one eligibility path, three entry points. Recording a grab is now a
  single transaction per release, so a multi-episode grab can no longer be
  half-recorded by a mid-write failure.
• Per-series search cadence lives in the store (last searched, backoff,
  next due) with an epoch guard so a reset landing mid-sweep — a series
  growing, or being re-monitored — beats the sweep's stale write.
• The recent-feed capability is a type assertion, not a wider interface,
  per the repo's optional-capability rule: an indexer that lacks it is a
  supported configuration, and decorators forward it only when the inner
  source really has one.