'no acceptable release' was ambiguous — couldn't tell if slskd returned nothing
or returned hits the quality profile rejected. now each item logs 'No search
results for X' vs 'N result(s) for X, none accepted — <reason>' (reason from the
top hit), and the summary tallies '… · N had no results, M rejected on quality'.
makes the slskd-finds-nothing-for-video situation legible.
bug: process_youtube_download wrote the ORGANISED dir (channel/Season YYYY) back
to the row's target_dir, but target_dir is supposed to be the youtube ROOT — and
plan_destination re-derives channel/season UNDER it. so any re-processing re-nested:
Channel/Season 2026/Channel/Season 2026/. that hit the 1-2 videos per channel that
got interrupted mid-download and re-queued by the orphan reaper.
fix: only record the organised filename for display; never write target_dir. the
root stays put so re-runs are idempotent. regression test runs it twice + asserts
no target_dir clobber + a stable (non-nested) dest dir.
the remaining badge bug: when a download finishes it removes its item from the
wishlist SERVER-SIDE, which fires no frontend 'changed' event — so the badge went
stale and only 'jumped' (e.g. down to the tv-only count) on the next navigation.
add a lightweight badge poll: refreshes the authoritative /wishlist/counts every
8s while downloads are active (uses the downloads page's _vdpgAnyActive), every
30s idle, paused when the tab is hidden. now the count tracks down live as wishlist
items get downloaded. frontend add/remove events still fire instantly as before.
the other half of the bug: the endpoint total was fixed, but setCounts() (movie/
episode load) AND setYtCounts() (youtube load) ALSO wrote the nav badge from their
OWN partial state. these load separately, and on the dashboard only one runs — so
whichever fired last overwrote the badge with a partial count (correct number,
then 'switches' to TV-only).
fix: the /wishlist/counts endpoint is the single source of truth for the grand
total; both setters now just call refreshBadge() to (re)sync from it instead of
computing a partial sum. regression test pins it.
bug: the wishlist badge read /wishlist/counts 'total', which was movies+episodes
only — YouTube videos (kind='video') are counted by a separate method and were
left out. so a wishlist of only youtube videos showed NO number; it only lit up
once tv episodes were added.
fix: the /wishlist/counts endpoint now folds youtube_wishlist_counts().video into
the total (+ exposes video/channel counts). the badge already live-refreshes on
wishlist-changed events (which youtube add/remove fire), so it now updates for
youtube too. db methods unchanged (their byte-identical contract is intact).
regression tests: mixed + youtube-only.
playlist cards on the watchlist tab now get the same hover cog → the per-channel
settings modal (custom show-name + quality override). the modal is kind-aware so
it reads 'Playlist settings' + 'the playlist's real name'.
no backend change needed: playlist videos are wishlisted with channel_id =
playlist_id, so the existing get/set_channel_settings + enqueue override lookup
already key off the playlist id. UI-only: cog on playlistCard + data-kind passed
to openChannelSettings(id, title, kind). test covers both cogs.
hover a channel on the watchlist Channels tab -> a settings cog appears (mirror
of the unfollow x) -> opens a self-contained modal:
- Show name (folder): overrides the $channel folder/show-name token; blank uses
the real channel name.
- 'Force a specific quality' toggle -> resolution/codec/container/60fps/HDR
override; off = use the global youtube quality from Settings.
VideoYoutube.openChannelSettings(id, title) builds the modal, loads via GET, saves
via POST /youtube/channel/<id>/settings. .vyt-cset-* CSS. string-contract test.
(note: 10 pre-existing soundcloud music-orchestrator test failures are unrelated
to this work — they fail on a clean HEAD too.)
storage + wiring + API for the channel settings modal (UI next):
- db.get/set_channel_settings(channel_id): per-channel {custom_name, quality}
in the settings KV store (no schema change); blanks clear the override.
- enqueue applies them: custom_name overrides the $channel folder token; a
quality override is stashed in the download row's search_ctx. the worker reads
it back (quality_override_from_download) and uses it instead of the global
youtube quality profile. global default still applies when no override.
- API GET/POST /youtube/channel/<id>/settings (GET also returns the global
default quality for the modal's 'using default' hint).
pure helpers (enqueue_ctx, quality_override_from_download) + db + api seam-tested.
closes the real gap: a restart kills the yt-dlp worker threads but leaves their
rows at 'downloading', which the pump counts as busy -> the queue wedges and they
never finish.
track live worker dl_ids in _active_worker_ids (added at worker start, dropped in
finally). requeue_orphaned_youtube() puts any 'downloading' youtube row with no
live worker back to 'queued' so the pump re-runs it. after a restart the set is
empty, so all stuck rows recover; during normal operation active ones are
protected (no false positives, no timestamp guessing). the hourly drain calls it
before pumping + logs the count. seam-tested.
the API returns system automations newest-created-first, so a re-seeded row
(e.g. the just-migrated movie/episode processors) jumps to the top and jumbles
the grouping. the video page now re-sorts its System list by an explicit order:
scans (fill) → processors (drain) → library scan/sync → maintenance. scoped to
the video page only (music ordering untouched); unknown/future actions fall to
the end keeping API order. string-contract test pins the scan<proc<maint order.
bug: renaming the action_types only changed code — a db already seeded under the
old names kept the stale rows (the seeder only CREATES missing ones, never removes
or renames). so on restart you got BOTH 'Auto-Download Movie Wishlist' (orphaned,
dead action) AND 'Auto-Process Movie Wishlist'.
add _fix_wishlist_processor_rename (runs in ensure_system_automations like the
other _fix_* migrations): deletes the orphaned video_download_movie/episode_wishlist
system rows (clearing is_system first, since delete_automation guards system rows)
and renames the youtube row in place (its action_type was unchanged). idempotent.
tested.
- 'Download X Wishlist' -> 'Process X Wishlist' everywhere (label + action_type):
matches the music side's 'process_wishlist', the already-process youtube action,
and reads right (movie/episode do search+pick+download, not just download).
action_types video_download_movie/episode_wishlist -> video_process_*.
- group the new automations as a two-stage pipeline in the seed list + builder
palette: Stage 1 SCANS (people/channels/playlists) fill the wishlist, Stage 2
PROCESSORS (movie/episode/youtube) drain it. icons/labels/drift test updated.
brand-new this session so no migration needed. 337 automation tests green.
the soulseek counterpart of the youtube drain — finally makes the people/airing
scans pay off. two automations (Auto-Download Movie / Episode Wishlist, hourly):
for each wished+released item, do a bounded blocking slskd search, pick the top
ACCEPTED release per the quality profile, and enqueue it exactly like a manual
grab (same add_video_download shape → the monitor finishes + organises it).
same standard as youtube: processes the WHOLE eligible wishlist (no total cap)
but searches a few at a time (max_concurrent, default 3, via a thread pool). a
busy guard skips the next hourly tick while a drain is still working. movies gate
on status='wanted' (skips monitored); episodes are all-wished. items already
downloading are skipped. quiet skip if the library folder isn't set.
reuses the real infra (build_query/slskd_search/_evaluate_hits/start_download/
download_monitor) — pure pick/select/record seam-tested; db queries added. 377
automation+video tests green.
playlists are followable but had no scan. new 'Scan Watchlist Playlists'
automation, sibling of the channel scan but a different rule: a playlist is a
curated finite set, so MIRROR it — wishlist every long-form video you don't have
plus any later additions (no forward-looking baseline, no last-N net).
playlist-as-show: videos wishlisted under the playlist's title, so the worker
files them as 'Playlist Name / Season YEAR / ... - date - title' (the ytdl-sub
tv_show_name-on-a-playlist convention). reuses all the channel plumbing — same
wishlist rows, same Download YouTube Wishlist drain, quality + org template.
seeded every 6h (Auto-Scan Watchlist Playlists); block + registration + icon +
drift test. dedup reuses wishlisted_video_ids_for_channel (parent_source_id=PL).
seam-tested; 944-test sweep green.
they were only builder blocks (not in the Active list). seed them like the
airing job so they appear + run out of the box:
- Auto-Scan Watchlist People — daily 03:00
- Auto-Scan Watchlist Channels — every 6h
- Auto-Download YouTube Wishlist — every 1h
scans no-op cleanly if you follow nothing. softened the download handler: an
unset youtube folder is now a quiet skip (status completed), not a per-run error,
so non-youtube users don't see a recurring failure.
ditch the per-run batch cap (a 200-video backlog would've taken weeks). now the
'Download YouTube Wishlist' automation queues the ENTIRE wishlist as 'queued'
rows, starts up to max_concurrent (default 3) right away, and each finished
download starts the next (one-out-one-in in the worker) so it all drains in a
controlled stream. the knob is 'max simultaneous downloads', not a total cap.
mirrors the music download worker's lesson (cap concurrency + space starts to
avoid yt-dlp 429s) but stays isolated on the video side:
- youtube_download: _pace() staggers fetch starts (3s); start_next_queued()
claims+spawns the next; run_youtube_download chains on finish.
- db.count_active_youtube_downloads() + claim_next_youtube_queued() (atomic,
race-safe). block field batch_size -> max_concurrent.
all seam-tested (pure select + pump); 680-test sweep green.
surface the youtube_template (built into organization.py) on Settings -> Library
Organization, next to the movie/episode templates: input + variable hints
($channel/$year/$date/$month/$day/$title/$videoid) + live preview, wired into
load/collect/save/reset. backend already round-trips it (load/save normalize).
new 'Download YouTube Wishlist' automation: pushes a polite batch of wished
youtube videos into the shared video_downloads queue + spawns the yt-dlp worker
per video. skips in-flight ones (no double-grab); big backlogs drain over
several scheduled runs (batch_size, default 3). needs the youtube library folder
set.
- download_monitor: SKIP source='youtube' rows (owned by their worker thread, no
slskd transfer to match) — surgical, slskd path untouched.
- db.youtube_wishlist_to_download(): flat newest-first list of wished videos with
channel/title/date/thumb for organising.
block + registration + icon/label + drift test. all seam-tested.
new core/video/youtube_download.py: fetch a wished youtube video end to end.
plans the organised dest (channel/year/date template), builds yt-dlp opts from
the quality profile (format_selection), runs the download (injectable factory),
then on success marks the video_downloads row completed + archives to history +
removes it from the wishlist; on failure archives failed and KEEPS the wish to
retry. orchestration is pure (yt-dlp run + all db writes injected) + seam-tested;
run_youtube_download binds the real seams for the worker thread.
flows through the SAME video_downloads queue as movies/tv (model B) so the
downloads page + history work for youtube for free.
add a 'youtube' scope to render_path: channel=show, season=upload year,
episode named '$channel - $date - $title'. rides the existing $token
engine (sanitised, dangling-separator tidy). undated videos fall back
cleanly (no empty 'Season ' / no stray ' - '). default template editable
like the movie/episode ones; per-channel override comes with the settings
modal later.
pure format_selection(profile) -> {format, format_sort, merge_output_format}.
caps to the resolution ceiling (falls back to uncapped so above-cap-only videos
still grab), ranks codec/res/fps/sdr as soft prefs (never excludes a stream).
the one piece the youtube downloader needs regardless of how the engine is wired.
exact yt-dlp tokens tunable on live yt-dlp; tests pin the shape.
two new video-side automation blocks that keep the wishlist fed:
- scan watchlist people: for each followed person, wishlist every un-owned
movie they acted in or directed (back catalog + upcoming). released ->
wanted, upcoming -> monitored (engine skips it til it's out, promotes on
release). grabs rich detail at add time (backdrop/cast/overview/etc +
provenance) into a new video_wishlist.detail_json col. fast re-runs skip
already-wishlisted + only promote.
- scan watchlist channels: for each followed youtube channel, wishlist new
long-form uploads (shorts excluded). forward-looking from follow time +
a last-N safety net (default 10). diffs against wishlisted/downloaded/
dismissed so it never dupes. pair with a 6h schedule trigger. scan-only;
fulfillment engine comes later.
both are pure handlers with injected seams + full seam tests. add_movie_to_
wishlist gains status + detail_json (promote-only upsert). no schema break,
music side untouched.
The mixed /trending/all chart is movie-heavy — TV was nearly absent (only
1 of the top 10 was a show). Split it like Netflix: 'Top 10 today' now holds
two ranked rails, Movies and TV Shows, from the dedicated /trending/movie/day
and /trending/tv/day charts (full 10 of each).
- client.trending(window, kind): single-type charts force the kind into
_disc_map (those endpoints omit media_type).
- engine.trending(window, kind): kind in the cache key.
- /discover/list: key=trending_movies_today / trending_tv_today, both treated
as single un-paged charts; inherit the same hide-owned + ranked rendering.
- frontend: one 'Top 10 today' group, two ranked shelves titled Movies / TV
Shows (group header carries the 'Top 10' framing).
Bug: with 'Hide owned' on, an owned title in the Top 10 (e.g. House of the
Dragon) had its card hidden by the global .vdsc-hide-owned rule while its big
rank numeral stayed — a blank gap with just a number. The ranked rail fetched
the true chart (ignoring hide_owned) precisely to keep the chart intact, which
collided with that CSS. Fix: ranked rail now honours hide_owned at fetch time
(owned dropped server-side, ranks stay contiguous 1-N), plus a CSS safety net
exempting ranked cards from the hide rule so a numeral can never be stranded.
When hide-owned is OFF the true chart still shows owned with the 'In Library'
ribbon.
Ordering: moved 'New & noteworthy' above the async collection/taste groups so
the two always-visible discovery rows (Top 10 + New) anchor the top of the
page, matching how streaming services surface new releases.
The page is now a long stack of rail groups (For you / Top 10 / New /
Trending / Mood / Studios / Genres / More). A sticky chip bar lets you
jump straight to any section — Netflix/Disney-style category nav. Chips
stay in lockstep with the groups: async-only groups (foryou/collection/
taste) reveal their chip when filled; pruned groups hide theirs. Lives
inside the shelves host so it auto-hides in the Browse-all grid view.
Foundation for the best-in-class discover build. All additive — existing callers unaffected:
- clients.discover() gains keywords (mood/theme), companies (studio), networks (TV), cast/crew
(people), min/max_runtime, certification (+country), vote_count_min override, and release_window
('last_30/90/365', computed date ranges) — each mapped to the right TMDB /discover param + gated
to movie/tv where TMDB requires.
- engine.discover_filter() threads them through + into the cache key.
- engine.trending(window) adds the real-time 'day' chart (cache keyed by window) for a Top 10 row.
- /discover/list parses the new query params and a key=trending_today shortcut.
py_compile + ruff clean. Frontend rails (Top 10, mood, studio, new&upcoming, quick-watches) next.
tests/video/__init__.py was created with the phase-1a gap-engine tests but never staged (those
commits added test files by explicit path). Every other test subfolder has a tracked __init__.py;
this keeps tests/video consistent and guards against import-file-mismatch under pytest's default
prepend import mode (the suite already has a real duplicate basename, test_selection.py).
The three async personalized loaders each did insertAdjacentHTML('afterbegin'), so 'Recommended
for you' / 'More like X' / gap rails landed in whatever order their fetch finished — the top of
the page reshuffled every load, and taste-based rails were scattered through generic ones.
Now Discover renders 6 authored groups in a fixed order, each with a header:
For you · Finish your collection · More of what you like · Trending & popular ·
Browse by genre · Hidden gems & more
Each async loader fills its own group's body (deterministic position) instead of racing to the
top; async-only groups (gaps) stay hidden until filled, and a group whose rails all drop out is
pruned. Extracted lazyShelfHtml/filledShelfHtml/shelfNav helpers; lazy-load, stagger, gen-guard,
see-all and scroll arrows all unchanged.
A heavy library with hide-owned on drained the old 8-page sequential fill before a rail filled,
leaving rows of 4-7 cards. Now the filtered-fill path pages up to 20 deep in concurrent waves of
5 (TTLCache + TMDB client are thread-safe), stopping as soon as it hits ~20 kept items or TMDB
runs out — so digging deep costs ~4 page-latencies instead of 20. Refactored the per-page filter
into consume(); trending + the no-filter path are unchanged in behaviour.
Toggling service/language chips calls reloadRails(), which clears the shelves synchronously
but the personalized loaders (foryou/gaps/morelike) fetch async and insertAdjacentHTML afterbegin.
Rapid re-toggles let superseded in-flight fetches prepend anyway -> N 'Recommended for you' rows
piled up. Each rebuild now bumps state.railGen and every loader bails if its captured gen is
stale before prepending; chip-driven rebuilds are also debounced (350ms) so multi-select
coalesces into one rebuild. The 'On your streaming services' rail itself was building fine -
it was just buried under the duplicates.
'My services' only builds the optional 'On your streaming services' rail, but sitting under
the 'Across Discover' header it read as a page-wide filter (user confusion: 'i thought it
affected the whole page because that's what it says'). Moved it to its own self-describing row
('pick what you subscribe to — adds a rail to your feed') below the bar. 'Across Discover' now
holds only the genuinely page-wide controls: Hide owned + Languages. Markup/CSS only — the
data-vdsc-myprov hook is unchanged.
Each rail showed a shimmer skeleton then revealed the whole row in a single fade — on a
big library with hide-owned on (which pages deep server-side) that reads as a long blank then
a pop. Cards now stagger in left-to-right via a per-card --i index (capped) + a 'backwards'
fill-mode keyframe (so the entrance animation releases and :hover transforms still work).
Applies to lazy rails (fillShelf) and the prepended personalized rows (staggerWithin).
The Browse-all grid silently inherited the global rail language preference, so users
couldn't ad-hoc browse foreign cinema without changing their homepage prefs. Added a
language chipset to the Browse panel (auto-wired via the generic chip handler -> state.sel.lang)
and the grid now always sends lang= : a real code filters, 'any' opts OUT of the rail
preference entirely. Route treats lang=any as 'no language filter'. Added a 'Browse the full
catalog' eyebrow so the panel reads as a self-contained search, parallel to 'Across Discover'.
Hide-owned and the saved 'My services' pref stay single/global by design (they apply page-wide);
the Browse panel's provider chips remain its own grid filter.
Hide-owned + language + my-services were inside the Browse-all filter panel, making them read
as grid-only when they actually affect every rail. Moved them into a labelled 'Across Discover'
strip above the Browse panel (JS finds them by data-attribute, so no logic change). Browse panel
now holds only its grid filters (kind/sort/genre/source/decade).
A saved streaming-services preference drives a personalized 'On your streaming services' rail:
- GET/POST /discover/providers-pref (TMDB provider ids); /discover/list OR-joins multiple
providers (comma->pipe) into with_watch_providers.
- A 📺 multi-select in the toolbar (Netflix/Prime/Disney+/Max/Apple TV+/Hulu/Paramount+/Peacock);
selecting services saves the preference and rebuilds the rails.
- The rail (high priority, after taste) appears once you've picked services, showing what's
streaming on yours.
- A ✕ 'Not interested' button on every un-owned Discover card (hover) — adds to the ignore list
and fades the card out instantly.
- A '🚫 Ignore List' button top-right of the hero opens a vibey glassmorphic modal: a header
explaining what it is, a search box to hide any movie/show directly (TMDB search), and a poster
grid of everything hidden with one-click 'Un-hide'. Empty state guides the user.
- Card button + modal both POST /discover/ignore; ignored titles vanish from every rail (via
_stamp_owned). Video-only, additive.
Add 'Hidden Gems' (movies) + 'Critically Acclaimed Shows' rails — vote_average.desc with the
backend's vote_count floor filtering out single-vote noise, and the language preference applied.
Slot them above the decade/foreign rails.
A 🌐 multi-select chip row in the Discover toolbar (EN/KO/JA/ES/FR/HI/DE/IT) to pick which
original languages appear in the general/curated rails. Loads the current preference from
/discover/languages, toggling a chip POSTs the new set and rebuilds the rails (never empty —
at least one stays on). Extracted reloadRails() (now shared by the hide-owned toggle + language
chips). Default EN, so the rails are English unless you opt more in.
The general/curated rails (Popular/Trending/Top Rated + genre/decade) pull TMDB's GLOBAL lists,
flooding feeds with foreign-language titles (Bollywood). Add a multi-language preference:
- _disc_map now carries original_language (+ popularity) on each item.
- discover_languages setting (default 'en'); /discover/list post-filters general/curated rails
to it (dropping known non-preferred-language titles) and pages deeper to keep rails full.
Rails with an explicit lang (the dedicated foreign rails) bypass the filter.
- GET/POST /discover/languages to read/set the preference.
- Removed the hardcoded lang=en on general rails (the setting drives it now).
Default 'en' immediately fixes the Bollywood flood; UI to pick languages next.
A single personalized wall aggregating TMDB recommendations across many of your owned titles
(random_owned_titles seeds), ranked by consensus — a title recommended by more of your library
ranks higher (ties by rating then popularity), owned + seed titles excluded.
- core/video/discovery_recs.py: pure blend_recommendations (dedup/consensus/exclude), 7 tests.
- /api/video/discover/foryou aggregates ~12 seeds' recommendations.
- loadForYou() prepends the 'Recommended for you' rail on top of the stack; re-runs on the
hide-owned toggle.
Two Discover UX issues:
- Foreign-language titles leaked into the general genre/decade rails. Added an
original-language filter (with_original_language) through client.discover -> discover_filter
-> /discover/list (?lang=); the genre/decade/'because you like' rails now pin lang=en, and a
handful of dedicated foreign rails (Korean/Japanese/Spanish/French/Hindi) house non-English.
- 'Hide owned' + a huge library = nearly-empty rails (a 2-page batch was mostly owned, then
CSS-hidden to almost nothing). /discover/list now takes hide_owned=1: it drops owned
server-side and pages DEEPER (up to 8) until a rail has ~24 un-owned. fillShelf passes
hide_owned when the toggle's on; toggling re-renders the rails (+ personalized rows) instead
of just CSS-hiding cards.
Already-matched movies predate the tmdb_collection_id column, so the collection gap rails
were empty (only newly-enriched movies got the id). Add a self-healing backfill: each
/discover/gaps load fills the franchise id for up to 20 owned movies missing it
(eng.movie_collection reuses the matcher's belongs_to_collection read), recording 0 for
movies with no franchise so they're not re-checked. The 'no-franchise' 0 is excluded from
the rails. Backfill is wrapped/isolated so it can never break the gap response. Over a few
Discover visits the whole library fills in and 'Complete the <franchise>' rails populate.
The #902 'paste cookies.txt' feature added a 'custom' sentinel value for
youtube.cookies_browser, but that feature wasn't merged to this branch — and ~7 call sites
(core/youtube_client.py x5, core/video/youtube.py, web_server.py) pass cookies_browser raw to
yt-dlp's cookiesfrombrowser, which rejects 'custom' ('ERROR: unsupported browser: custom') and
broke YouTube download/enrichment. Sanitize 'custom' -> '' (no browser) at every site:
youtube_client reads via a walrus filter, the other two guard the condition. 'custom' now
means 'no browser cookies' here (the cookiefile feature isn't on this branch). Latent on dev
too — only _youtube_cookie_opts was fixed there.
The new index on movies.tmdb_collection_id was in video_schema.sql, which executescript runs
BEFORE _ensure_columns adds the column on existing DBs — so 'CREATE INDEX ... ON
movies(tmdb_collection_id)' failed with 'no such column' and the whole video DB init aborted
(500s on every /api/video/* call). Moved the index to _POST_INDEXES (runs after the ALTERs),
matching the pattern the code comments already prescribe. The CREATE TABLE columns stay (fresh
DBs) + the ALTER migration stays (existing DBs); only the index moved.
loadGaps() fetches /discover/gaps and prepends the gap rails ('Complete the <franchise>',
'More from <director/creator>') above the rail stack, mirroring loadMoreLike. Cards are the
standard un-owned TMDB cards — already actionable (VideoGet add-to-watchlist/get), so a
missing franchise entry or director film is one click from your queue. Best-effort/additive.
- video_database: owned_movie_tmdb_ids (diff set), owned_movie_collections (franchises you've
started, most-invested first), top_owned_people (directors/creators you own the most).
- engine.collection(id): cached + owned-annotated franchise film list (person_detail already
gives owned-annotated filmography).
- /api/video/discover/gaps: builds 'Complete the <franchise>' rails (collection_gaps) + 'More
from <person>' rails (filmography_gaps, movies, vote-filtered) — the 'what am I missing' section.
All additive; gap diffs are the pure tested core.
core/video/discovery_gaps.py — two pure diffs powering the 'what am I missing' rails:
collection_gaps (franchise entries you don't own, in collection order) and
filmography_gaps (a person's titles you don't own, deduped, kind/vote-filtered, ranked
by popularity). No I/O — the API wires owned-ids/collection-items/person-credits in.
9 tests.
Data layer for the 'complete your collections' gap engine. People/credits/genres are already
normalized + indexed, so the only missing signal was franchise membership:
- movies: + tmdb_collection_id (indexed) + tmdb_collection_name (schema + _COLUMN_MIGRATIONS,
SCHEMA_VERSION 17->18, _ENRICH_META_COLS whitelist so enrichment_apply backfills them)
- enrichment match() reads belongs_to_collection (a standard movie-detail field, no extra call)
and writes the id/name into the match metadata.
Additive + backfill-only (COALESCE), nothing existing rewired. (also noqa'd a pre-existing
OMDb S110 in the touched file to keep ruff clean.)
