feat(reco): radio + similar with metadata fallback (§6.5)

POST /radio + /radio/next (stateless infinite feed: seed track / from-likes,
exploration mix, client-passed exclude_ids) and GET /tracks|artists/{id}/similar,
replacing the stubs. Recommender port abstracts the (future) ML service —
NullRecommender is wired now so RecommendationService always uses its metadata
heuristics (genre/artist similarity, random exploration filler), never a hard ML
dependency. Adds TrackRepository.list_similar/sample_playable + Artist.list_similar,
reason codes for the client, RemoteRecommender skeleton (TODO: ML contract).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/api/schemas/radio.py

@@ -0,0 +1,43 @@
+"""Radio + similarity response schemas (§6.5)."""
+
+import uuid
+
+from pydantic import BaseModel, Field
+
+from app.api.schemas.artist import ArtistOut
+from app.api.schemas.track import TrackOut
+
+
+class RadioRequest(BaseModel):
+    """Start or continue a radio. ``seed_track_id`` seeds from a track;
+    ``from_likes`` seeds from the caller's likes. ``exclude_ids`` are already-
+    queued tracks to skip (the client drives the infinite feed). ``exploration``
+    biases familiar↔new."""
+
+    seed_track_id: uuid.UUID | None = None
+    from_likes: bool = False
+    exploration: float = Field(default=0.25, ge=0.0, le=1.0)
+    count: int = Field(default=20, ge=1, le=50)
+    exclude_ids: list[uuid.UUID] = Field(default_factory=list)
+
+
+class RadioTrackOut(BaseModel):
+    track: TrackOut
+    # Short code the client localizes: ml | similar | from_likes | discover.
+    reason: str
+
+
+class RadioResponse(BaseModel):
+    # Where the picks came from: "ml" or "metadata" (fallback).
+    source: str
+    tracks: list[RadioTrackOut]
+
+
+class SimilarTracksOut(BaseModel):
+    source: str
+    tracks: list[TrackOut]
+
+
+class SimilarArtistsOut(BaseModel):
+    source: str
+    artists: list[ArtistOut]