A delivered gap is not a missing translation. The delivery arrives and says
this target language does not voice this utterance, with a reason - the
translation renders a different paragraph, say. That is a different problem
with a different fix, and the reviewer needs to be able to tell them apart.
The reason was being thrown away. The row is stored with an empty text and
a review_reason; the resolver skipped every row with empty text, so the gap
was dropped and the review lane said "Chinese has not been translated yet".
That is a false statement about a translation that was delivered and
rejected.
So the resolver now keeps a not-voiced row as a gap with its reason, the
chapter page renders it as a gap rather than as missing text, and the
language page lists the delivered gaps under "Needs attention" with their
reasons. A gap row is also no longer counted as translated text, so the
coverage the reviewer sees is the coverage the book has.
4 new tests. Suite: 913 passed, 8 failed (the same 8 pre-existing failures).
Also adds scripts/probe_cross_language_tts.py: the open prerequisite for any
target-language audio lane is whether Qwen3-TTS can speak Chinese with an
English voice reference. The probe answers it by transcription rather than
by listening - Whisper transcribes each clip and reports what language it
heard, and the English-reference clip reading Chinese is compared against
the same clip reading Chinese with English phonology, so the only variable
is the language parameter.
ff73c99f8d