four fixes from the review (and a self-correction):
1) close the connection. reconcile_unverified_history_from_tracks opened a connection with no
finally/close. runs once per boot so GC reclaimed it, but now it's consistent + robust.
2) scope the tracks scan to the review queue. it built lookup dicts from EVERY verified/
human_verified track (~350k on a large library) on every boot while anything is unverified
(the normal state). now it loads the stuck rows first and skips verified tracks whose path
AND basename can't match any queued row, so dicts stay proportional to the queue, not the
library. behaviour identical (all 13 PR reconcile tests still pass).
3) close the title-less basename collision. a title-less history row fell back to filename-only
matching with no ambiguity check, so a generic name like "01 - Intro.flac" could heal a
DIFFERENT song to verified. now a title-less basename heal only fires when that basename is
unique among verified tracks; unique-basename rows still heal (recall preserved).
4) "Clean orphaned" protects force_imported rows (deliberate user decision, keep for human
approval) without weakening the mount-down safety gate. CRUCIAL self-correction: filtering
them out BEFORE the orphan check (my first cut) shrank the checked count below the threshold
and would have let a few unverified orphans be deleted during a mount outage. instead,
find_orphan_history_ids now takes a deletable predicate: protected rows still count toward
checked / all-missing (gate stays strong) but never enter the orphan_ids delete set.
3 new regression tests (title-less collision; deletable protects from delete; protected rows
still count toward the gate). 936 verification/acoustid/history/downloads tests green. builds
on nick2000713's #938.
clicking Download Discography → Add all to wishlist added ~1 track every 15-30s. trace: the
endpoint's per-track library-ownership check (track_already_owned → check_track_exists) ran the
LEGACY path — firing search_tracks for every title-variation × artist-variation, per track. on
a large library and an artist you own NOTHING of, STRATEGY-1 (indexed LIKE) always missed and
fell through to the fuzzy fallback (full-table scan), ~10-15 scans/track = the 15-30s. metadata
fetch was never the bottleneck (deezer returns each album in ~1s).
fix: pre-fetch the artist's owned tracks ONCE (get_candidate_albums_for_artist →
get_candidate_tracks_for_albums) and pass candidate_tracks to check_track_exists's batched
in-memory path — the same path the discography backfill job + completion-stream already use.
pass an EMPTY list (not None) when nothing is owned so the owns-nothing case still takes the
fast path → instant. per-track cost drops from ~20s to ~1ms.
safe: track_already_owned's only real caller is this endpoint; the new param is optional
(None = unchanged legacy behaviour for any other caller). normal ownership still detected (same
artist-variation breadth); the one divergence is a track owned ONLY via a compilation → a
harmless redundant wishlist add, which is the endpoint's explicitly-accepted failure mode and
already how the backfill job behaves. 4 new tests; 1134 discog/metadata/wishlist tests green.
the dashboard orb canvas kept falling back to ~1fps on firefox after the page settled. root
cause: firefox throttles requestAnimationFrame to ~1fps for a canvas it heuristically deems
occluded. the WAA keepalive only delayed the heuristic; it re-fired over time.
real fix: on firefox, drive tick() with a setInterval(~60fps) instead of self-scheduling rAF —
setInterval is not subject to the canvas-occlusion rAF throttle, so the orbs stay at full
framerate indefinitely. chrome is untouched (keeps vsync-aligned rAF). same render workload
(idle still drops to 20fps via the existing sleep skip); background tabs still parked by the
visibilitychange handler → stopLoop clears the interval. kept the keepalive (harmless).
#936 added fragmented-row grouping. two follow-ups found in review:
1) BUG: an excluded canonical sibling (a row pinned to the same canonical edition whose tracks
fail the strict fragment match) was emitted with canonical_items but no owned set, so
_build_missing_tracks flagged the ENTIRE tracklist as missing — including tracks the row
already owns (e.g. a 3-track fragment of a 12-track edition reported 12 missing, not 9).
now compute that sibling's own owned slots from its local tracks (shared
_owned_reference_for_tracks helper, same logic as the anchor) so it reports only what it's
actually missing and the count stays internally consistent.
2) PERF: _build_candidate_groups rescanned all albums for every canonical group — O(G*N), which
degrades badly once many editions are pinned (fine today at G=1, latent at scale). invert to
a single O(N) pass that assigns each row to its unique group; identical 'exactly one match'
semantics. also added a Stop/Pause check in _prepare_work_items, which now front-loads the
canonical lookups + matching.
3 new tests (excluded-sibling reports only its missing tracks; ambiguous candidate stays
independent; unambiguous candidate joins) — 17 completeness + 123 repair tests green.
root cause: the library stores album/artist art as media-server RELATIVE paths (Plex
/library/metadata/.., Jellyfin /Items/.., Navidrome /rest/..), which don't render in a browser
<img>. normal wishlist items carry Spotify CDN urls so they show fine, but LIBRARY-sourced
items — dead-file re-downloads and preview-clip re-fetches — carry the raw relative path, so
their album art came up blank. and the nebula only had artist photos for WATCHLISTED artists,
so non-watchlist orbs showed initials.
fix on READ in the wishlist tracks endpoint (so it also repairs items already in the wishlist,
no re-run needed), using the library data we already have:
- normalize each track's album.images url that needs it — relative/internal only, via the
canonical normalize_image_url; CDN urls are left untouched so already-rendering items can't
regress.
- build an artist-name -> normalized library-photo map and return it; the nebula seeds its
artist-image map from it (every wishlist artist), with curated watchlist photos overriding.
8 tests (predicate: relative/internal fixed, CDN untouched; album normalize in-place; artist
map build/skip-empty/idempotent/graceful). 237 wishlist+repair+JS tests green, ruff clean.
two field reports:
1) FIX-ALL skipped these findings — bulk_fix_findings() has a hardcoded fixable_types
allowlist that didn't include 'short_preview_track'. added it, so select-all/fix-all and
the per-page bulk-fix now cover this tool.
2) RE-WISHLISTED ITEM WAS ART-LESS — the payload pulled album art from the library thumb,
which is empty for un-enriched HiFi previews, so the wishlist orb showed initials (no album
OR artist image, since the orb falls back to album art). now the duration lookup also
captures the metadata source's CDN art from raw_data (spotify album.images / itunes
artworkUrl, upscaled) and stores it on the finding; the fix prefers that over the empty thumb.
3 new tests (art capture from spotify raw_data, itunes artwork upscale, fix uses finding art);
8 job tests + repair suite green.
The reconcile heals rows whose file is still in the library; it deliberately
leaves ORPHANS — history rows whose file is gone (deleted / replaced /
re-downloaded elsewhere). Those can never be healed (no file left to confirm)
and linger in the Unverified list forever. This adds an explicit, user-initiated
cleanup for them.
- core/downloads/orphan_history.py: pure, tested rule. A row is an orphan when
its file resolves nowhere; flags `suspicious` when EVERY reviewed file is
unreachable (the mount-down signature) so the caller refuses rather than
mass-delete a healthy log during an outage.
- POST /api/verification/clean-orphans (admin-only): runs it against
_resolve_history_audio_path (raw path -> prefix-swap resolver -> tracks-table
title fallback), refuses on the suspicious signature, and deletes only history
ROWS — never a file (the files are already gone).
- UI: "🧹 Clean orphaned" button in the Unverified bulk-actions row, with a
confirm dialog spelling out that it removes log rows only and refuses if the
library looks offline.
NEVER automatic / never at boot — a filesystem check during a mount outage would
otherwise wipe good history. 5 pure-rule tests + safety-gate coverage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LWJk7EuM7YktQeNyqQwTZY
Second-pass audit of the startup reconcile found a real correctness bug in it:
the basename fallback had NO title guard, so a shared track-number filename
("01 - Intro.flac" in different albums) would heal the WRONG song — marking an
actually-unverified file 'verified' and silently dropping it from the review
queue. This is the exact collision the scan-time matcher (history_match.py)
guards against; the reconcile now mirrors it.
- basename match now requires the history row's title and the candidate track's
title to agree (alphanumeric-lowercase), only when BOTH are present (legacy
rows without a title still fall back to filename-only, like the matcher).
- exact-path matches stay unguarded (same path = same file, unambiguous).
- cheap early-out: skip the tracks scan entirely when no 'unverified' rows exist
(keeps the every-boot cost ~nil on healthy libraries).
3 new tests (collision must-not-heal, titles-agree heals, missing-title falls
back). 8 reconcile tests green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LWJk7EuM7YktQeNyqQwTZY
each preview-clip finding now renders a dedicated detail card: File Length vs Real Length
(e.g. 28s vs 200s) and a Play button so the user can listen to the clip and confirm it's a
busted ~30s preview before approving the delete + re-download. reuses the existing
_renderPlayButton/playFindingTrack path that dead_file/orphan findings already use (the
finding carries file_path + title/artist/album, everything it needs).
two field-reported issues: clicking the job sat at 'Starting…' with Scanned: 0 forever and
never showed the current track.
1) HANG — spotify get_track_details() defaults to allow_fallback=True, which scrapes the
configured metadata source when the official API isn't authed (HiFi users). that scrape is
slow and blocked the scan loop on the first track. now pass allow_fallback=False (official
only — fast, returns None cleanly) and fall through to iTunes/MusicBrainz.
2) NO LIVE UPDATE — progress was only pushed every 5 tracks via update_progress, never the
current item. now report_progress() every track (phase + 'artist — title' + scanned/total)
plus a start phase, so the UI moves and shows what it's checking.
also made the test track ids INTEGER to match production (tracks.id is INTEGER PRIMARY KEY),
exercising the real str(id) finding -> WHERE id=? round-trip. 5 tests green.
HiFi (and occasionally other) downloads sometimes deliver a ~30s preview clip instead of the
full song; it lands in the library looking real. new repair job scans short tracks (duration
<= 30s, configurable), looks up the EXPECTED length from the track's metadata source
(spotify/itunes/mb get_track_details), and flags any whose real length is much longer than the
file (default: >= 30s longer) as a preview clip.
approving the finding (repair_worker._fix_short_preview_track) deletes the preview file (path
resolved via _resolve_file_path like the other delete tools), drops the DB row so the track
goes missing, and re-adds it to the wishlist with the full payload (mirrors _fix_dead_file)
so the real version downloads. scan ONLY creates findings — nothing destructive without user
approval, like every other tool.
conservative: genuine short tracks (source agrees they're short) and tracks whose length can't
be verified are skipped, never flagged. registered the job + finding-type label/fix-button in
the UI. 5 tests (scan flag/skip/scope + fix delete+remove+wishlist); 89 repair tests green.
the duration-agreement check used abs() drift, so a file running LONGER than the metadata
(a remaster with a longer outro, an extended cut) was rejected the same as a truncated one —
e.g. A-Ha 'Take on Me' remaster at 228.5s vs 225.0s expected, quarantined for +3.5s.
but a longer file is the OPPOSITE of truncated. make the auto tolerance asymmetric: keep the
tight 3s/5s bound for SHORTER files (the truncation case the check exists for), allow up to
15s in the LONGER direction for version/master differences. a wrong song still trips it (off
by far more than 15s), and a user-pinned tolerance is honoured symmetrically. direction-aware
rejection message too. 4 new tests; 274 integrity/import tests green.
the growth-triggered collects were firing but RSS still climbed to 2.2GB before snapping to
1.2GB — because gc.collect() freed the python objects but glibc hoarded the memory rather than
returning it to the OS, so RSS stayed at the high-water mark. add malloc_trim(0) after each
collect to hand freed arenas back to the OS, so incremental collects genuinely lower RSS and
the sawtooth caps near floor+200MB instead of overshooting. best-effort (skipped on musl/non-
linux). also tightened the growth trigger 250->200MB.
the 60s timer overshot: browsing piled plexapi cyclic garbage faster than once-a-minute caught
it, so RSS hit ~2.2GB before a sweep (then dropped to 1.2GB). switch to polling RSS cheaply
(every 8s) and collecting as soon as it grows +250MB since the last sweep — so it fires DURING
a heavy browse and caps the peak near floor+250MB instead of running to 2GB+. keeps a 120s
backstop for slow idle accumulation.
measured the 'resource hungry' / lockup issue: browsing every page grows RSS ~300MB -> 1.8GB
and it stays. it's not a leak — it's deferred cyclic collection. plexapi parses Plex responses
into XML Element trees whose nodes reference each other in cycles; Python's generational GC
leaves them in gen2 and sweeps it rarely, so ~227k Element objects pile up. forcing gc.collect()
reclaimed ~700MB instantly (1.8GB -> 1.1GB live), confirming.
add a daemon that runs a full gc.collect() every 60s so the cyclic garbage is reclaimed on a
cadence instead of climbing into lock-up. full collect is ~tens of ms; once a minute is
negligible. this is the root of the reporter's 2GB + ramonskie's spike too.
tracemalloc's continuous tracing locks up a loaded app, so add a one-shot gc-based memory
breakdown: top object types by total size AND by count, plus the biggest individual containers
(>1MB). a runaway 'count' points at an unbounded cache; a big bytes/str total points at blob
retention. lets us pinpoint the RSS growth (300MB -> 1.8GB after browsing) without tracing.
the Memory Usage stat showed only global system memory (psutil.virtual_memory().percent).
add the process's own resident set size (RSS) — the real 'how much RAM SoulSync uses' number —
formatted MB under 1GB, GB above. headline stays the system %, subtitle now reads 'SoulSync ·
612 MB' instead of the generic 'Current usage'. graceful fallback if psutil errors / older
backend. useful context after the recent RAM-footprint discussions.
Firefox re-rasterizes blur()/backdrop-filter every composite where Chrome caches it, so the
always-visible shell glass (sidebar header + aura orbs, hero/header buttons) was ~half of
Firefox's idle GPU. gate behind @supports(-moz-appearance:none) so it's Firefox-only: hide the
two blur(28px) sidebar orbs + the dash-card blobs, and drop backdrop-filter on the sidebar
header and hero/header buttons (each keeps its tint, just unfrosted). measured ~20-25% -> ~10-13%
on Firefox, every page (sidebar is always visible). chrome is untouched — the block doesn't
exist there, full frost intact.
removing the always-on dash-card blob animation (for the Chrome GPU win) incidentally let
Firefox start throttling the worker-orb canvas's compositing to ~1fps after a header hover
re-layerizes the dashboard — Chrome never throttles it. re-add the 'keep the compositor warm'
effect cheaply: a 2px, ~invisible element running an infinite transform-only animation (zero
paint). gated behind CSS.supports('-moz-appearance') so it's Firefox-only; Chrome never gets
it. confirmed fix in Firefox/Zen.
the .dash-card cursor blobs are 16 large blur(48px)/blur(18px) layers. chrome caches them
once; firefox re-rasterizes blur on every composite, so they're a big chunk of idle dashboard
GPU on firefox. they're purely decorative and reduce-effects already hides them. gate behind
@supports(-moz-appearance:none) so it's firefox-only — chrome keeps the full cursor glow,
this block doesn't exist there.
same antipattern as the dash-card blobs: .sidebar::before/::after are blur(28px) and the
orb keyframes animated transform: scale() infinitely → the GPU re-blurred them every frame,
on every page (the sidebar is always visible). keep the translate drift + opacity (both
compositor-only, the blur layer just moves), remove the scale. same look, no per-frame reblur.
each .dash-card renders two accent-blob pseudo-elements — ::before is 1280x1280 blur(48px),
::after 540x540 blur(18px) + mix-blend-mode:screen — and both ran an INFINITE scale-pulse
animation. scaling a blurred element re-rasterizes the blur every frame; with 8 cards × 2
blobs that's 16 huge blurred layers re-blurring at 60fps whether or not the user touches
anything. that's the dashboard's whole-screen repaint / ~36% idle GPU.
remove the infinite pulse (the dashBlob*Pulse animations). the blob still follows the cursor
via --blob-x/y; it just no longer 'breathes' at idle, so when nothing's moving there's nothing
to repaint. trimmed will-change to the props that actually change (left/top).
the full-page particle canvas runs a continuous requestAnimationFrame loop behind every
page — real GPU cost, and multiple users hit GPU strain until they found the toggle. flip
the default to off; the eye candy is opt-in now.
- init.js: runtime flag defaults false unless localStorage is explicitly 'true'
- settings.js: config read is now '=== true' (default off) instead of '!== false'
- index.html: checkbox no longer 'checked' by default; hint reworded
existing users who explicitly enabled it (localStorage/config 'true') keep it on; the
existing '!== false' runtime guards still work since the flag is now always set explicitly.
The Quarantine sub-view was reworked into rich cards (artwork, source line,
row-click to expand an inline detail panel, consistent action cluster) but the
Unverified sub-view was left on the generic download-row layout that opened a
modal on click. Bring it to parity:
- dedicated _verifUnverifiedRowHtml / _verifUnverifiedRows renderer, used via a
sibling branch in _adlRender (mirrors the quarantine sub-view branch).
- row click toggles an inline details panel (why flagged, download source,
quality, file, downloaded-at), open-state keyed by stable id so it survives
the 2 s poll re-render — same pattern as verifQuarInspect.
- reuses the existing verif-quar-* / verif-actions / adl-row CSS (no new styles)
and the existing play/compare/audit/approve/delete handlers.
- NO grouping: each unverified import is its own track (grouping only makes
sense for the quarantine alternates), per design intent.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LWJk7EuM7YktQeNyqQwTZY
Complements the AcoustID-scan-time heal: re-links library_history rows still
showing 'unverified' to the verified/human_verified status their file already
carries in the tracks table — matching exact path AND basename, so a file that
moved (media-server import / reorganize) heals even though the stored history
path is frozen. Upgrade-only and non-destructive (no deletes, no bulk
migration).
Why this is needed on top of the scan-time fix:
- It clears the EXISTING backlog (e.g. 5551 rows) on the next restart with NO
re-fingerprinting and no AcoustID API calls — the file's status is already in
tracks from the prior scan; this just propagates it to the frozen history row.
- It covers human_verified files, which the AcoustID scan skips entirely
(file_verif_status == 'human_verified' returns early), so their stale history
rows would otherwise never heal.
Runs once on DB init (cheap, idempotent). 5 real-sqlite tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LWJk7EuM7YktQeNyqQwTZY
the AcoustID scanner matched library_history rows by EXACT file_path, but that path is
frozen at import time while the file moves afterward (media-server import / reorganize) —
so tracks.file_path (what the scan reads) no longer equals it. two failures resulted, both
introduced in 37ea6604: verified status never reached the history row (verified tracks kept
showing 'unverified'), and a fresh acoustid_scan row was INSERTed every run (5551 rows for
3675 songs).
- new pure, tested matcher (core/downloads/history_match.py): exact path → filename guarded
by title; prefers a real download row over a synthetic scan row.
- _persist_status now HEALS the matched row's path + status (so future scans match cleanly),
DELETES synthetic acoustid_scan duplicates by exact path (collision-free, never a real row),
and inserts only when the file genuinely has no row.
- a full AcoustID job now self-cleans existing duplicates — no destructive bulk migration.
8 matcher + 4 real-DB heal/dedup/insert tests; existing scanner tests updated to the new
seam (heal vs insert). 1076 acoustid/verification/download tests green.
save_watchlist_scan_run had a single caller — the manual scan endpoint. the automatic/
scheduled path (process_watchlist_scan_automatically) ran the full scan but never wrote a
history row, so nightly scans never showed up in the History modal — only manual ones.
- new shared helper persist_scan_run(database, state, ...) extracts the run from the
finished watchlist_scan_state and writes one history row
- the automatic path now stamps scan_run_id/scan_track_events and calls it
- the manual path is refactored onto the same helper so the two can't drift apart again
- history is global (no profile filter), so the all-profiles nightly scan records one
aggregate row (profile_id None → 1, never NULL)
tests: 4 new persist_scan_run seam tests (real DB) + 2 new auto-scan integration tests
proving the auto path actually records (completed + cancelled, exactly once). 420
watchlist/automation tests green.
a pasted track link IS resolved + searched, but the 'bubble the exact track to the top'
step read getattr(t,'id') — and TrackResult has no top-level id (the source id lives in
_source_metadata['track_id']). so the bubble was a silent no-op: the linked track sat buried
among fuzzy text-search lookalikes and the user saw unrelated tracks. qobuz made it worse —
_qobuz_to_track_result never stamped _source_metadata at all, so the track had no id to match.
- stamp _source_metadata={'source':'qobuz','track_id':...} on qobuz TrackResults (mirrors tidal)
- extract the bubble into pure, tested helpers (linked_track_id / bubble_linked_track_first)
that read _source_metadata['track_id'] — fixes it for tidal too, str/int-safe, stable no-op
19 track-link tests (+6 new) + 87 qobuz/download tests green.
_normalize_album_for_match stripped ANY trailing '- clause', so a real distinguishing
subtitle ('Clair Obscur: Expedition 33 - Nos vies en Lumière (Bonus Edition)') collapsed to
the same name as the OST → _albums_likely_match treated them as one album → the watchlist
marked unowned tracks of one edition as owned via the other and under-wishlisted.
- strip a trailing '- ...' clause ONLY when every token in it is an edition/format
qualifier (+ connectors / year-ordinal): '- Single', '- Acoustic Version', '- 2011
Remaster' still collapse, but real subtitles ('- Nos vies en Lumière', '- Volume 2',
'- Live in Berlin') are kept. Avoids the inverse regression (a same-album pair splitting
into a redownload loop), which a naive narrow strip would have caused.
- drop the loose substring shortcut + raise the fuzzy floor 0.6→0.85; genuine drift already
collapses to an EXACT match, so the looseness only ever produced false fuses.
blast radius: _albums_likely_match has exactly one caller (the allow-duplicates skip).
48 album-match tests pass (qualifier-suffix merges + edition-subtitle splits) + 219 watchlist.
the file faked spotipy + config.settings in sys.modules at import time with no teardown.
the fake config.settings had no ConfigManager, so depending on collection order it leaked
into tests/test_config_save_retry and intermittently failed the full suite. the real
modules import fine in the test env (spotipy is installed, config.settings has both
ConfigManager + config_manager), so the stubs were pure liability — removed them. album
tests still pass (10), the album+config combo that errored now passes (17), 573 repair/
config/canonical tests green.
#929 added 'musicbrainz' to library_reorganize._ALBUM_ID_COLUMNS but not to the
canonical layer, breaking the equality invariant test (canonical reads exactly what
reorganize reads). add musicbrainz to CANONICAL_ALBUM_SOURCES, and move it from the
'can't pin' param group to the 'pins' group in the manual-lock tests — now consistent,
and forward-compatible with pinning a deliberately-matched MB edition. inert at runtime
today (mb isn't in the manual source selector, so should_pin is never called with it).
540 canonical/reorganize tests green.
test_personalized_playlists_id_gate asserted the OLD (wrong) deezer thresholds (>=100000, raw-rank assumption) — the same bug fixed in c033656f. The discovery pool synthesizes deezer popularity to 0-100, so the test now asserts (60, 50). This is the CI failure from running the full suite (my -k subset missed it).
results comes from asyncio.gather over to_search, so they're always equal length — strict=True asserts the invariant and satisfies ruff B905. Carried in with #896.
The discovery pool synthesizes deezer popularity onto a 0-100 score (base 45 + bonuses, capped at 100), but _get_popularity_thresholds had deezer on the raw-rank scale (500000/100000). So Popular Picks' 'popularity >= 500000' matched nothing — empty for every deezer-primary user — while Hidden Gems' '< 100000' caught the whole pool. Deezer thresholds now sit on the 0-100 scale (60/50, like Spotify's 60/40). Tested.
Every library track was stored with disc_number=1 because the Jellyfin/Plex/Navidrome scan parsed the track number but never the disc field. Multi-disc albums collapsed onto disc 1, so disc-2+ tracks were mis-filed (shown under disc 1) and flagged 'missing' — the frontend title-fallback band-aid couldn't recover it (breaks on iTunes title mismatches).
Now the shared insert_or_update_media_track reads the disc number (Jellyfin .discNumber=ParentIndexNumber, Navidrome .discNumber, Plex .parentIndex), floors to >=1, and stores it in the INSERT + UPDATE. The disc_number column is ensured on init (it was only added by a migration that doesn't run on fresh installs, so the new INSERT would have hard-failed for new users). The enhanced album view already carries disc_number through (SELECT * -> dict), so the display fixes itself once the column is populated — a re-scan backfills existing libraries. Seam-tested across Jellyfin/Navidrome/Plex shapes + the floor-to-1 + re-scan-update cases.
Opens to the same category-card landing the Discovery Pool uses, with two cards: 'guesses to review' (unverified wing-it) and 'resolved manually' (ones you've Fixed) — click to drill in, Back to return. Previously it jumped straight to a single list.
To populate the resolved list, the /fix endpoint now stamps was_wing_it on the rewritten extra_data (the wing_it_fallback flag is otherwise lost on fix), and get_wing_it_pool gained a resolved flag + the stats return both counts. Fixing/re-matching from either card refreshes in place. Seam test updated for both states.
Wing It auto-matches tracks to the server library on a best-effort guess; those tracks are flagged wing_it_fallback in extra_data and count as 'discovered', so the Discovery Pool hides them — there was no way to see or audit the guesses. New 'Wing It Pool' button (next to Discovery Pool on the Mirrored Playlists tab) opens a modal listing them with a per-playlist filter + search; 'Fix Match' reuses the Discovery Pool's fix flow (/api/discovery-pool/fix), and a manual match drops the track from the pool on refresh.
No new table or provider hooks needed — the wing-it flag is already persisted, so this is a pure query (get_wing_it_pool / get_wing_it_pool_stats, cloning the failed-pool LIKE pattern) + a /api/wing-it-pool endpoint + a cloned modal. Found 81 wing-it tracks on a real library. Seam-tested (include unverified / exclude manual-matched / scope by playlist+profile).
The card is an <a> link and the shell's capture-phase link handler navigated to artist-detail before the grid's bubble-phase badge handler could preventDefault — so clicking the watchlist eye or a source badge opened the detail page (and the badge's own link too). The shell handler now bails when the click lands on an in-card control (.source-card-icon or [data-no-card-nav]), letting the badge do only its own thing.
