UI UX improvements

CONTEXT.md

@@ -67,3 +67,16 @@ tracks that *are* present, which streamrip writes into every file. An album is:
   **Missing disc**.
 - **Unknown** — completeness cannot be determined, because no present track has
   readable tags to reveal the expected total.
+
+### Source
+A single music service (Qobuz, Tidal, Deezer, SoundCloud, ...) and everything
+that is specific to it: how to build a share URL from an id and read an id back
+out of one (**url** / parse, both directions), how to fetch its **album art**,
+and how to read **metadata** from one of its URLs. A Source owns its
+credentials/keys too (Qobuz, for instance, reads its app_id and auth token from
+the **Streamrip database**'s config). Each Source reaches the network only
+through the injectable `http_get` seam, mirroring the `run_rip` and
+`read_audio_tags` seams, so an adapter can be exercised with canned JSON and no
+real network. Sources are migrated behind this single interface one at a time;
+a not-yet-migrated source keeps its quirks inline (e.g. in `construct_url` and
+`extract_metadata_from_url`) until its slice lands.

Dockerfile

@@ -15,11 +15,19 @@ RUN groupadd -g 1000 appuser && \
 # Set working directory
 WORKDIR /app
 
-# Install Python dependencies
+# Install Python dependencies. streamrip is installed from the upstream `dev`
+# branch (not the PyPI release, which lags behind): PyPI's latest is 2.1.0
+# while dev/host run 2.2.0+, and the config schema version is tied to the
+# streamrip version — a mismatch makes host and container fight over the shared
+# self-mounted config.toml (each rewrites it to its own schema on every run).
+# Tracking dev keeps the container aligned with the host build.
+# Note: `@dev` is a moving target, so rebuilds are not reproducible; Docker
+# layer caching means this only re-pulls when this layer is invalidated
+# (e.g. `--no-cache` or an earlier-layer change).
 RUN pip install --no-cache-dir \
     flask \
     flask-cors \
-    streamrip \
+    "git+https://github.com/nathom/streamrip.git@dev" \
     gunicorn \
     gevent
 
@@ -38,5 +46,12 @@ USER 1000:1000
 # Expose port
 EXPOSE 5000
 
-# Run with aggressive worker recycling
-CMD ["gunicorn", "--bind", "0.0.0.0:5000", "--worker-class", "gevent", "--workers", "2", "--timeout", "60", "app:app"]
+# Single worker only: Active/History state, the SSE subscriber list, and the
+# download worker threads all live in-process (ADR-0002). A second gunicorn
+# worker is a separate process with its own copy of all three, so the browser's
+# SSE stream and the request that starts a download can land on different
+# workers — the download runs but never shows up in Active/History. The gevent
+# worker class serves many concurrent SSE clients in this one process via
+# greenlets, and download concurrency is handled internally by
+# MAX_CONCURRENT_DOWNLOADS worker threads, so one worker loses nothing.
+CMD ["gunicorn", "--bind", "0.0.0.0:5000", "--worker-class", "gevent", "--workers", "1", "--timeout", "60", "app:app"]