Milestone 4: yt-dlp provider and weighted source scheduler

Add a second playback source and a weighted scheduler mixing it with
Navidrome:

- Scheduler picks a provider by SOURCE_WEIGHTS, falling through to the
  others when one has nothing ready, so no source can stall playback.
- YtdlpProvider reads a hand-maintained config/urls.txt; container URLs
  (playlist/album/label/artist) are flat-extracted and one entry is
  drawn at random, honouring the anti-repeat window. Adds Track.source_url.
- YtdlpFetcher downloads bestaudio via the yt-dlp library, reusing the
  atomic hidden-temp-then-rename pattern; Liquidsoap decodes the result.
- Queue now dispatches to a fetcher registry keyed by backend.
- Sweep orphaned download temp files on daemon startup (leftovers from a
  killed container otherwise pile up and trip the stream fallback).

Verified end-to-end: yt-dlp opus decoded and served as 192 kbps MP3, and
the 3:1 default mix observed in play history.

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

README.md

@@ -58,16 +58,27 @@ cp .env.example .env
 If the Navidrome variables are left empty, the source is simply disabled and
 the stream plays whatever is already in `cache/` (the milestone-1/2 behaviour).
 
+For the yt-dlp source, list the URLs to draw from in `config/urls.txt` (copy
+`config/urls.txt.example`). Each line is either a direct track URL or a
+container URL (playlist, album, label, artist page) from which one track is
+picked at random. The relative mix between sources is set by
+`RADIEO_WEIGHT_NAVIDROME` / `RADIEO_WEIGHT_YTDLP` (a weight of 0 disables a
+source); the file being absent also disables yt-dlp.
+
 ## Current status
 
-**Milestone 3 — Navidrome provider: done.**
+**Milestone 4 — yt-dlp provider: done.**
 
-- `ingest` pulls tracks from an OpenSubsonic playlist (Navidrome), downloading
-  them into the shared cache ahead of playback (prefetch buffer).
+- Two playback sources feed a weighted scheduler: a Navidrome/OpenSubsonic
+  playlist and a hand-maintained list of yt-dlp URLs (`config/urls.txt`).
+  Container URLs (playlist/album/label/artist) are expanded and one track is
+  drawn at random, honouring the anti-repeat window.
+- Each source has its own fetcher (Subsonic stream / yt-dlp download); files are
+  cached ahead of playback (prefetch buffer) and decoded by Liquidsoap.
 - Play history and LRU retention are tracked in a SQLite database under
   `state/`: only the N most recently played files are kept on disk
-  (`RADIEO_RETENTION_KEEP`, default 20); anti-repeat avoids replaying a track
-  seen among the last plays.
+  (`RADIEO_RETENTION_KEEP`, default 20). Orphaned download temp files are swept
+  on startup.
 - `GET /next` returns the next track as an annotated Liquidsoap URI with real
   title/artist metadata (or an empty body when nothing is ready).
 - `stream` (Liquidsoap v2.4.5) pulls via `request.dynamic` and falls back to the
@@ -75,7 +86,7 @@ the stream plays whatever is already in `cache/` (the milestone-1/2 behaviour).
 - HTTP stream served at `http://localhost:8000/radio.mp3` (MP3, 192 kbps),
   multiple simultaneous listeners supported.
 
-The yt-dlp and ListenBrainz sources come next.
+The ListenBrainz suggestion feed comes next.
 
 ## Roadmap
 
@@ -84,8 +95,8 @@ The yt-dlp and ListenBrainz sources come next.
    switches to a `request.dynamic` source with the cache as fallback.
 3. ✅ **Navidrome provider** — play from an OpenSubsonic playlist, with caching,
    LRU retention and play history.
-4. **yt-dlp provider** — fetch tracks from a maintained URL/artist list; weighted
-   mixing between sources.
+4. ✅ **yt-dlp provider** — fetch tracks from a maintained URL/artist list;
+   weighted mixing between sources.
 5. **Canonicalizer** — ListenBrainz MBID lookup for source-agnostic
    de-duplication.
 6. **ListenBrainz provider** — parse the RSS suggestions feed and resolve each