feat(tn/en): opt-in roman-numeral list markers (roman_enumerators)

[Alex-Wengg]

Counterpart of FluidInference/FluidAudio#974 (fixes FluidAudio#972 there).

Kokoro read (i), (ii), (iv) as letters. NeMo's roman grammar is uppercase-only and keyword-anchored, so the compiled FST leaves every list-marker form untouched (verified: (i) … (ii) … (iv), (IV), II. Scope, ii) noise all pass through; only World War II converts).

What — NormalizeOptions::roman_enumerators (default false, same shape as disable_bare_second) + tn::en::roman::spell_enumerators, a port of FluidAudio's EnglishTextNormalizer pre-pass:

• (ii) anywhere unless glued to a letter (f(x), café(i) stay)
• ii) / ii. only at line start or after ; : , . + whitespace; dot form only before whitespace (i.e. stays)
• I V X only, single case, strict form, 1–39 (keeps mix, cd, xl out)
• markers that aren't self-evident list items (uppercase, lone v/x, all-x) need a second enumerator in the text — (IV) intravenous, checkbox (x), xx. sign-offs, v. Madison, I. M. Pei stand alone; an uppercase outline I. … II. … III. converts whole

Plumbing — tn_normalize_sentence_lang_with_options (rules), fst::normalize_lang{,_with_options} (FST), FFI nemo_tn_fst_with_options + nemo_tn_normalize_sentence_lang_with_options, WASM tnNormalizeSentenceLangWithOptions, Swift wrapper overload, both headers.

Parity — flag off is byte-identical to before on every path: FST parity suite green, nemo_tn_fst untouched, FFI test asserts off == plain.

Reviewer notes

• Ran: cargo test, cargo test --features ffi,fst-engine (incl. fst_parity), cargo build --target wasm32-unknown-unknown --features wasm, cargo fmt --check.
• FluidAudio keeps its Swift pre-pass regardless (needed for the trait-off build); this flag is for the WASM/npm and other non-Swift consumers.

🤖 Generated with Claude Code