fix(worker): cancel abandoned slskd transfers + cap download fall-through

Two compounding causes of 'many downloads of one album from different sources' in
slskd: (1) the slskd client raised on timeout/error WITHOUT cancelling the transfer
it enqueued, so as the pipeline fell through to other peers each left a transfer
running; (2) the incomplete-download fall-through was uncapped, grinding through
every candidate of a popular album (~200 Soulseek peers), amplified now that Qobuz
pacing pushes most jobs to Soulseek.

- SlskdClient.download now cancels its enqueued transfers (DELETE) on any abandon.
- The download loop caps at _MAX_DOWNLOAD_ATTEMPTS (6) candidates.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

worker/lyra_worker/pipeline.py

@@ -24,6 +24,11 @@ _AUDIO_EXT = {".flac", ".mp3", ".m4a", ".opus", ".ogg", ".aac", ".wav"}
 # track COUNT right. Generous, so edition/encoding differences don't false-positive; bonus
 # tracks (over-long) are always fine.
 _DURATION_MIN_RATIO = 0.85
+# Cap how many ranked candidates a single job will actually download+verify before giving up.
+# Without this, the incomplete-download fall-through can grind through dozens of sources (a popular
+# album returns ~200 Soulseek candidates), each a slow attempt that also leaves an abandoned slskd
+# transfer. The best candidates rank first, so 6 attempts is plenty.
+_MAX_DOWNLOAD_ATTEMPTS = 6
 
 
 def _measure_staged_duration_s(staging: str) -> float | None:
@@ -397,10 +402,14 @@ def run_pipeline(
     _poller.start()
     try:
         try:
+            attempts = 0
             for candidate in ranked:
                 adapter = by_source.get(candidate.source)
                 if adapter is None:
                     continue
+                if attempts >= _MAX_DOWNLOAD_ATTEMPTS:
+                    break  # don't grind through every peer/source of a popular album
+                attempts += 1
                 _mark_chosen(conn, job_id, candidate.source_ref)  # reflect the source now in flight
                 shutil.rmtree(staging, ignore_errors=True)  # clean slate per attempt
                 _reports.clear()  # this attempt hasn't reported yet → poller estimates until it does