Italian (if_sara/im_nicola, lang_code='i') audio always truncated to fixed 0.3s regardless of input length

Bug description

Using lang_code='i' (Italian, espeak-ng fallback) with either Italian voice (if_sara or im_nicola), every synthesized clip is truncated to a fixed length of exactly 7200 samples (0.3s at 24kHz), regardless of input text length — from a single word up to full sentences of 100+ characters. Confirmed reproducible with kokoro 0.7.4, 0.7.9, and 0.7.16 (same result on all three), on macOS (Apple Silicon, M1).

This is not "long text runs fast" — the output sample count is identical across wildly different inputs, and each is missing most of the intended sentence.

Repro

from kokoro import KPipeline
pipeline = KPipeline(lang_code='i')

for text in [
    'Ciao.',
    'Ciao Marco come stai oggi',
    'Questa è una frase molto più lunga per vedere se la durata dell audio cambia in base alla lunghezza del testo che sto scrivendo qui adesso',
]:
    for _g, _p, audio in pipeline(text, voice='if_sara'):
        print(len(text), 'chars ->', len(audio), 'samples', round(len(audio) / 24000, 3), 'sec')

Output on my machine:

5 chars -> 7200 samples 0.3 sec
25 chars -> 7200 samples 0.3 sec
138 chars -> 7200 samples 0.3 sec

Same result whether text is a single string or pre-split into separate \n-joined sentences (each sentence chunk independently comes back as exactly 7200 samples).

What I ruled out

• Not a G2P/vocab issue: phonemization via espeak-ng produces correct, full-length IPA (verified against espeak-ng --voices=it), and every phoneme in the string maps to a valid model.vocab entry (KModel.vocab.get(p) — checked manually, zero Nones, so nothing is silently filtered before reaching the model).
• Not a permissions/environment issue: same exact 7200-sample result across 3 separate installs (fresh venvs), 3 different kokoro versions, with misaki at both its bundled version and 0.9.4 installed manually.
• English works correctly — same machine, same pipeline pattern, lang_code='a', voice='af_heart': a comparable-length English sentence produces ~94800 samples (3.95s), scaling normally with text length.
• Audio content is not silence — it's real (non-zero) waveform data, just cut off very early.

Environment

• macOS, Apple Silicon (M1)
• Python 3.12
• kokoro 0.7.4 / 0.7.9 / 0.7.16 (all three tested, same result)
• espeak-ng 1.52.0 (Homebrew), Italian voice pack present (espeak-ng --voices=it lists it)

Expected

Audio duration should scale with input text length for the Italian (espeak fallback) voices, similar to the native English/Japanese/Chinese pipelines.

Related but not identical: #301 (Spanish em_alex returns silence on Windows) — different symptom (empty vs. fixed-length truncation) and different platform, so filing separately rather than piling on there.