`<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>
f7942f3689