fix(gp8): confine registry asset matching to the declared directory (#1011)

`<EmbeddedFilePath>` is matched on filename stem so a format variant of the
same recording can win — an `.ogg` beside the declared `.mp3` is copied out
losslessly rather than transcoded. But the search spanned every directory in
the archive, so an unrelated file that merely shared the stem could stand in
for the declared asset: exactly the substitution the registry lookup added in
#1007 exists to prevent.

Candidates are now confined to the registry path's own directory. A genuinely
absent asset still falls through to the legacy stem match and then the first
audio asset, as documented.

Found by an adversarial pass over #1007 rather than a report — no known file
triggers it, since GP8 writes embedded audio to Content/Assets/ and that is
the only directory scanned. It needs a hand-edited archive to reach.

Both tests fail on main and pass here; their ZIP ordering is deliberate, so
the fall-through target differs from the decoy (otherwise fixed and unfixed
code return the same file and the tests prove nothing).


Claude-Session: https://claude.ai/code/session_01SFDokqh2H6mEjk1Kgbi6JW

Signed-off-by: ChrisBeWithYou <christian.a.cowan@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

tests/test_gp_audio_sync.py

@@ -375,3 +375,51 @@ def test_legacy_stem_match_still_works_without_a_registry():
     stem, path = _resolve_audio_asset(zf)
     assert stem == "abc-123"
     assert path == "Content/Assets/abc-123.ogg"
+
+
+def test_a_same_stem_file_in_another_directory_cannot_stand_in():
+    """The registry names a PATH, not just a name.
+
+    Resolution matches on stem so a format variant of the same recording can
+    win, but an unrelated file that merely shares the stem must not satisfy
+    the declaration — that substitution is what the registry lookup exists to
+    prevent. The declared asset is genuinely absent here, so the right answer
+    is the documented fall-through, not the decoy.
+
+    ZIP order matters to this test: `real.ogg` is written FIRST so the
+    fall-through target differs from the decoy. Otherwise both the fixed and
+    unfixed code return the same file and the test proves nothing.
+    """
+    buf = io.BytesIO()
+    with zipfile.ZipFile(buf, "w") as zf:
+        zf.writestr("Content/score.gpif", _gpif_bytes(
+            "0", {"0": "Content/Audio/track.ogg"}))
+        zf.writestr("Content/Assets/real.ogg", b"fake")     # fall-through target
+        zf.writestr("Content/Assets/track.ogg", b"decoy")   # shares the stem only
+    buf.seek(0)
+    _, path = _resolve_audio_asset(zipfile.ZipFile(buf))
+    assert path == "Content/Assets/real.ogg", (
+        "a same-stem file in a directory the registry never named must not "
+        "satisfy the declaration"
+    )
+
+
+def test_the_declared_directory_still_resolves_its_own_format_variants():
+    """The directory constraint must not cost us the OGG preference.
+
+    The shallower decoy is written FIRST, so unfixed code (which searches
+    every directory) picks it and this test fails.
+    """
+    buf = io.BytesIO()
+    with zipfile.ZipFile(buf, "w") as zf:
+        zf.writestr("Content/score.gpif", _gpif_bytes(
+            "0", {"0": "Content/Assets/nested/take.mp3"}))
+        zf.writestr("Content/Assets/take.ogg", b"decoy-one-level-up")
+        zf.writestr("Content/Assets/nested/take.mp3", b"declared")
+        zf.writestr("Content/Assets/nested/take.ogg", b"same-take-lossless")
+    buf.seek(0)
+    stem, path = _resolve_audio_asset(zipfile.ZipFile(buf))
+    assert path == "Content/Assets/nested/take.ogg", (
+        "the OGG variant in the DECLARED directory wins over a shallower decoy"
+    )
+    assert stem == "take"