feat(discovery): acquire recommendations that keep their names

Acquisition ran yt-dlp without --embed-metadata, so every download
arrived untagged. The scanner then stored the video id as the title and
"Unknown Artist" as the artist, the vetted-candidate tag check rejected
the mismatch, and all 18 acquired tracks were hidden and retired.

- Pass --embed-metadata so downloads carry real tags.
- Let a scan take fallback title/artist from the candidate, for sources
  that still ship untagged files.
- Install Deno alongside yt-dlp: YouTube guards some formats with a JS
  challenge yt-dlp must execute, and no other runtime is enabled.
- Dedupe candidates by artist and title. The (source, external_id) key
  misses the same song reaching us under two Deezer release ids.

Also carries the in-flight discovery work this builds on: the
Recommendations page replacing Discover, the discovery source service,
and the acquisition spec tests.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

backend/src/db/migrations.ts

@@ -759,4 +759,64 @@ export const MIGRATIONS: Migration[] = [
       ON CONFLICT (session_id) DO NOTHING;
     `,
   },
+  {
+    // Two external candidate strategies, kept as separate source_trust keys so
+    // meta-learning can compare their retention independently: "artists you
+    // already play just released something" is a much stronger prior than
+    // "Last.fm thinks this sounds similar", and the trust values say so.
+    id: '20260806_external_discovery_sources',
+    sql: `
+      INSERT INTO source_trust (key, trust, description) VALUES
+        ('new_release', 0.60,
+         'New release by an artist already played from the local library.'),
+        ('similar_recommendation', 0.45,
+         'External similarity (Last.fm) seeded from local play history.')
+      ON CONFLICT (key) DO NOTHING;
+    `,
+  },
+  {
+    // play_history is the only durable record of what was listened to and when,
+    // and it had two holes that only show up when you try to read a year back:
+    //
+    //  1. ON DELETE CASCADE meant the gated cleanup sweep silently erased the
+    //     plays of every file it removed. A play happened; deleting the file
+    //     later does not un-happen it. The FK becomes SET NULL and the track's
+    //     identity is denormalised onto the row so it stays readable.
+    //  2. No duration, so listening time was only ever inferable from the
+    //     track's current duration — itself gone once the file is.
+    id: '20260806_play_history_durable_facts',
+    sql: `
+      ALTER TABLE play_history ADD COLUMN IF NOT EXISTS listened_ms INTEGER;
+      ALTER TABLE play_history ADD COLUMN IF NOT EXISTS track_title TEXT;
+      ALTER TABLE play_history ADD COLUMN IF NOT EXISTS track_artist TEXT;
+
+      DO $$
+      DECLARE fk_name TEXT;
+      BEGIN
+        SELECT con.conname INTO fk_name
+          FROM pg_constraint con
+          JOIN pg_class rel ON rel.oid = con.conrelid
+          JOIN pg_attribute att ON att.attrelid = rel.oid AND att.attnum = con.conkey[1]
+         WHERE rel.relname = 'play_history'
+           AND con.contype = 'f'
+           AND att.attname = 'track_id'
+           AND con.confdeltype = 'c'
+         LIMIT 1;
+        IF fk_name IS NOT NULL THEN
+          EXECUTE format('ALTER TABLE play_history DROP CONSTRAINT %I', fk_name);
+          ALTER TABLE play_history
+            ADD CONSTRAINT play_history_track_id_fkey
+            FOREIGN KEY (track_id) REFERENCES tracks(id) ON DELETE SET NULL;
+        END IF;
+      END $$;
+
+      -- Backfill identity for rows written before the columns existed. Rows whose
+      -- track was already cascade-deleted are unrecoverable; this at least stops
+      -- the bleeding from here on.
+      UPDATE play_history ph
+         SET track_title = t.title, track_artist = t.artist
+        FROM tracks t
+       WHERE t.id = ph.track_id AND ph.track_title IS NULL;
+    `,
+  },
 ];

backend/src/db/schema.sql

@@ -302,12 +302,20 @@ CREATE TABLE IF NOT EXISTS recommendation_batch_track (
 -- "Success-Driven Center" rule (a completed play moves the active batch's center)
 -- and the feedback learning loop. Created with IF NOT EXISTS so it can be
 -- self-provisioned on databases that predate this schema change.
+-- A play is a historical fact: it stays true after the file is gone. Hence
+-- ON DELETE SET NULL rather than CASCADE, plus the denormalised title/artist so
+-- a row whose track was hard-deleted is still readable. listened_ms is what was
+-- actually heard, recorded at play time; tracks.duration is not a substitute
+-- because it disappears with the track.
 CREATE TABLE IF NOT EXISTS play_history (
     id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
     user_id UUID NOT NULL,
-    track_id UUID REFERENCES tracks(id) ON DELETE CASCADE,
+    track_id UUID REFERENCES tracks(id) ON DELETE SET NULL,
     batch_id UUID REFERENCES recommendation_batch(id) ON DELETE SET NULL,
     completed BOOLEAN NOT NULL DEFAULT false,
+    listened_ms INTEGER,
+    track_title TEXT,
+    track_artist TEXT,
     played_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
 );