fix(worker): retry transient MusicBrainz failures with backoff

The matching stage makes up to three musicbrainzngs calls per job, none of
which retried. A dropped TLS connection (surfacing as NetworkError,
"<urlopen error [SSL: UNEXPECTED_EOF_WHILE_READING]>") or a 503 rate-limit
(ResponseError) failed the entire pipeline for the job, which then re-queued
and re-ran from scratch — ~36% of jobs churned this way over a 6h window.

Wrap the three MB calls in _with_retry (4 attempts, 1s/2s/4s exponential
backoff), catching only the two transient musicbrainzngs error types so real
bugs still propagate. musicbrainzngs stays lazily imported to keep module
import offline.

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

worker/tests/test_musicbrainz.py

@@ -1,4 +1,7 @@
-from lyra_worker._musicbrainz import _best_release_group
+import musicbrainzngs
+import pytest
+
+from lyra_worker._musicbrainz import _best_release_group, _with_retry
 
 # Offline unit tests for the pure release-group selection logic. The live resolve()
 # path (network) is covered by test_musicbrainz_live.py.
@@ -60,3 +63,73 @@ def test_prefers_credited_artist_over_same_titled_other_artist_album():
     ]
     best = _best_release_group("Lake Street Dive", "Fun Machine", groups)
     assert best["artist-credit"][0]["artist"]["name"] == "Lake Street Dive"
+
+
+# --- _with_retry: transient MusicBrainz failures (SSL EOF, 503 rate-limit) ---
+
+
+def test_retry_returns_value_without_backoff_when_first_call_succeeds():
+    sleeps: list[float] = []
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        return "ok"
+
+    assert _with_retry(fn, sleep=sleeps.append) == "ok"
+    assert calls["n"] == 1
+    assert sleeps == []  # no backoff when nothing failed
+
+
+def test_retry_recovers_from_transient_network_error():
+    # The real symptom: <urlopen error [SSL: UNEXPECTED_EOF_WHILE_READING]> arrives
+    # as musicbrainzngs.NetworkError. A couple of retries clears it.
+    sleeps: list[float] = []
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        if calls["n"] < 3:
+            raise musicbrainzngs.NetworkError("SSL: UNEXPECTED_EOF_WHILE_READING")
+        return "recovered"
+
+    assert _with_retry(fn, sleep=sleeps.append) == "recovered"
+    assert calls["n"] == 3
+    assert sleeps == [1.0, 2.0]  # exponential backoff before each retry
+
+
+def test_retry_recovers_from_transient_response_error():
+    # MusicBrainz's own 503 rate-limit surfaces as ResponseError.
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        if calls["n"] < 2:
+            raise musicbrainzngs.ResponseError("503 Service Unavailable")
+        return "ok"
+
+    assert _with_retry(fn, sleep=lambda _: None) == "ok"
+    assert calls["n"] == 2
+
+
+def test_retry_reraises_after_exhausting_attempts():
+    sleeps: list[float] = []
+
+    def fn():
+        raise musicbrainzngs.NetworkError("still down")
+
+    with pytest.raises(musicbrainzngs.NetworkError):
+        _with_retry(fn, attempts=4, sleep=sleeps.append)
+    assert sleeps == [1.0, 2.0, 4.0]  # 3 backoffs across 4 attempts
+
+
+def test_retry_does_not_swallow_non_transient_errors():
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        raise ValueError("bug, not a network blip")
+
+    with pytest.raises(ValueError):
+        _with_retry(fn, sleep=lambda _: None)
+    assert calls["n"] == 1  # non-transient errors propagate immediately, no retry