Narrator speaks the translation language; characters keep their own accent

Specification change. Two different rules, and they need different
mechanisms.

Narrator: accent = translation language. Its reference was generated by
VoiceDesign from a description that hardcoded '{accent} English', baking an
American-English timbre into the WAV; cloning that into German gives German
phonology from an American voice. Ten per-language references now exist
(narrator_EN.wav ... narrator_ZH.wav), each designed from the narrator's
timbre plus 'Speaking {language} as a native {language} speaker'.

Characters: keep the accent on the speaker page. Their accents were already
encoded in their existing refs - what destroyed them was the language
parameter forced onto every utterance, which made every voice adopt the
target language's prosody. That is why Deal Hunter came out French-native.
Phonology and accent are now separate controls: language carries phonology,
an explicit voice_description carries accent character. Deal Hunter gets
'with a strong Australian accent' in every language.

Measured A/B, French line from Deal Hunter: with and without the Australian
instruction the waveforms differ by 152% of signal amplitude and duration
moves 4.24s -> 4.08s, so the instruction is reaching the model. Whether it is
convincingly Australian is an ear test, not something measurable here.

Re-rendered all 120 clips under the new spec, 0 failures. The previous
render is kept in Lang_test_preSpec for A/B. 24 speakers carry a non-default
accent; the narrator is deliberately excluded from that set.

Suite: 721 passed, 8 pre-existing.

specs/29-accent-and-voice-spec.md

@@ -0,0 +1,83 @@
+# Accent and voice specification for translated books
+
+Status: specification, agreed 2026-10-13. Implemented by
+`scripts/generate_narrator_refs.py` + `scripts/render_lang_test.py`.
+
+## The rule
+
+Two rules, and they are different:
+
+1. **The narrator speaks with the accent of the translation language.** An
+   Italian book is narrated by an Italian-sounding narrator. A German book by a
+   German one. Not an American reading Italian.
+
+2. **Every other character keeps the accent defined on the speaker page.**
+   Deal Hunter stays Australian. Duchess Victoria stays Canadian. Xu Ming stays
+   Hong Kong. This holds in every target language.
+
+The asymmetry is deliberate. The narrator is a narrator in *each* language and
+is heard far more than any single character, so a foreign accent on the
+narrating voice is the thing listeners notice. Characters are defined
+individually on the speaker page, and that definition is the contract.
+
+## Why this needed two different mechanisms
+
+**The narrator** needed a new reference voice per language. Its reference was
+generated by the VoiceDesign model from a description that hardcoded
+`"{accent} English"`, which baked an American-English timbre into the WAV.
+Cloning that into German gives German phonology from an American voice. Passing
+the target language fixes phonology but not identity.
+
+So the narrator now has one reference per language, generated by VoiceDesign from
+the narrator's timbre plus "Speaking {language} as a native {language} speaker".
+`narrator_EN.wav` … `narrator_ZH.wav`, ten in all, in
+`<project>/.umbra/voices/`.
+
+**Characters needed no new references.** Their accents are already encoded in
+their existing reference WAVs. What was destroying them was the `language`
+parameter: forced onto every utterance, it made every voice adopt the target
+language's prosody, which is why Deal Hunter came out French-native.
+
+Phonology and accent are separate controls, and the fix uses them separately:
+
+| | language parameter | accent carried by |
+|---|---|---|
+| Narrator | target language | its per-language reference |
+| Characters | target language | existing reference + explicit accent instruction |
+
+The accent instruction is derived from the speaker page (`accent_for()` in the
+renderer) and passed as `voice_description` alongside the clone. Measured A/B on
+one French line from Deal Hunter: with and without her Australian instruction,
+waveforms differ by 152% of signal amplitude and duration moves 4.24s -> 4.08s.
+
+## Deal Hunter's current accents
+
+24 speakers have a non-default accent. The narrator is deliberately *not* in
+this list — its accent is the translation language.
+
+    australian  Deal Hunter (strong)        canadian   17 speakers incl.
+    nigerian    Adonai Maul, Dealer                   Kainda, Nanny,
+    indian      Alec, Interrogator                     Duchess Victoria
+    kenyan      Warbots                    hong_kong  Trader Wakam, Xu Ming
+
+Any character whose accent is `none` simply gets no instruction and speaks the
+target language plainly.
+
+## What this does not fix
+
+- **The ICL prompt path is still unusable.** One `prompt_slug` request takes the
+  server from 5.4 GB to 19.1 GB and never returns. The render path uses
+  `x_vector_only_mode=True`, which is stable. Cached prompts are effectively
+  bypassed.
+- **Accent is verified by measurement, not by ear.** The instruction demonstrably
+  changes the output. Whether "strong Australian" is *convincingly* Australian
+  in French is a listening test for the reviewer, not something measurable here.
+- **One pass per language.** Languages are rendered sequentially so the system
+  can reset between them; not all fifteen at once.
+
+## Division of labour with the translation node
+
+The translation node produces the text and the print paperback, including the
+full analysis needed for correct translation. This system turns that text into
+audio and knows nothing about translation. The renderer consumes
+`(index, speaker, text)` per utterance and emits WAV.