Fix thumbnail backoff and watchlist fallback matching

README.md

@@ -2,7 +2,7 @@
 
 A local web UI for a system-wide `ani-cli` install.
 
-Current version: `0.26.0`
+Current version: `0.29.0`
 
 ## Project Layout
 
@@ -11,6 +11,8 @@ ani-cli-web/
   app.py          Public entrypoint, runtime wiring, watchlist domain logic, and CLI startup.
   app_support.py  Shared runtime/state/config helpers plus queue and library file utilities.
   queue_jobs.py   Download queue and background watchlist-refresh workers.
+  watchlist_identity.py Safer watchlist identity resolution for queued download jobs.
+  title_matching.py Shared anime-title normalization and matching helpers.
   http_handler.py HTTP route handling, request parsing, and remote-access guard.
   web_templates.py Shared Search, Config, and Watchlist page markup aggregation.
   search_page.py  Search page template.
@@ -70,7 +72,10 @@ The app currently includes:
 - The watchlist page now also avoids re-triggering the same unresolved thumbnail warm-up request on every silent poll cycle while a show's thumbnail state is unchanged.
 - Background `Refresh All` execution with status polling, parallel per-show refresh work, and no inline thumbnail downloads during the bulk job.
 - Search and Watchlist polling now use a shared serial browser polling helper so slow responses do not cause overlapping poll requests to pile up in the background.
+- That shared browser polling helper now also tears itself down on page unload and uses chained `setTimeout` scheduling instead of a bare repeating interval.
 - Controlled shutdown and runtime reset now cancel active download workers when waiting for teardown, reducing the chance of orphaned background downloads outliving the web process.
+- Runtime reset now also refuses to clear the active global runtime when background workers fail to stop cleanly, preventing stale workers from outliving the runtime object that created them.
+- Watchlist refresh networking now uses a shorter dedicated timeout, which helps per-show and bulk watchlist shutdown paths settle more predictably when upstream metadata services are slow.
 - Persistent watchlist refresh history that survives restarts and marks interrupted bulk refresh jobs clearly.
 - Local thumbnail caching plus manual thumbnail upload when automatic cover lookup fails.
 - A refreshed glassy dark UI inspired by the Tsuki frontend design language.
@@ -80,6 +85,8 @@ The app currently includes:
 - Search-page queue polling also guards against stale out-of-order refresh responses now.
 - Queue log writes are batched during active downloads to reduce SQLite churn on noisy jobs.
 - Queue persistence now stores core job metadata and logs in structured SQLite columns instead of treating the whole job as one large JSON blob.
+- Failed thumbnail refresh attempts now record a retry timestamp, so unresolved covers back off instead of immediately requerying upstream sources on every later warm-up.
+- Download-to-watchlist fallback matching now keeps season and part identity intact while still tolerating equivalent season notation such as `2nd Season` vs `Season 2`.
 - Project-local runtime state under `.ani-cli-web/`.
 
 ## Search Page
@@ -189,6 +196,9 @@ Current behavior:
 - If the anime is already on the watchlist, a successful download marks it as downloaded and keeps its current category.
 - If the anime is not yet on the watchlist, `ani-cli-web` creates a new watchlist entry in `Finished` after the download succeeds.
 - Those add and download-sync paths now queue a background refresh for episode counts, AnimeSchedule metadata, and thumbnails instead of waiting for all upstream fetches before returning.
+- If a queued job ever reaches the older search-based fallback path for `show_id` resolution, the app now only accepts that fallback when the queued title still matches the returned result confidently, preferring to fail safe over marking the wrong anime as downloaded.
+- That fallback now also preserves season and part distinctions, so `Season 2` downloads do not silently sync against a different season with the same base title.
+- When that safer fallback still cannot resolve a `show_id`, the failure now keeps the specific reason so logs and error messages are easier to diagnose.
 - Existing watchlist entries from older versions are migrated into the `Watching` category automatically.
 
 ### Thumbnail Behavior
@@ -207,7 +217,11 @@ Current thumbnail behavior:
 - If a season-labeled title fails as-is, lookup retries with simplified base-title queries.
 - AniDB is used as a fallback when AnimeSchedule does not resolve a cover.
 - AniDB fallback reuses the cached parsed title list directly to keep repeated lookups lighter.
-- Failed fetch attempts no longer refresh the retry window, so a bad lookup does not get stuck in an immediate repeat-404 state.
+- Failed automatic fetch attempts now start the retry window, so bad lookups back off instead of hammering AnimeSchedule and AniDB on every warm-up pass.
+
+### Watchlist Pagination
+
+Watchlist listing responses now clamp out-of-range page requests to the last available page before querying rows, so the returned `page` number and `items` stay consistent even when the browser or API client asks for a page past the end.
 
 Per-card thumbnail tools: