Albanian audio in one call

You are integrating the Zani text-to-speech API (natural Albanian voices) into this project.

## What Zani is
Zani turns text into natural Albanian speech (dialects: standard, kosovo, gheg, tosk). It is a REST API: JSON in, JSON out, Bearer auth.
Base URL: https://www.tryzani.com
Full docs: https://www.tryzani.com/en/docs

## Before you write code
1. Inspect the project: language, framework, where server-side code lives, how secrets/env vars are handled, and any existing audio/TTS code. Follow the project's conventions.
2. If it is not obvious, ask me where audio should be generated and what to do with it (play in the UI, save a file, serve to users).
3. Ask me to create an API key at https://www.tryzani.com/app/developers (it starts with "zn_live_") and store it in the environment variable ZANI_API_KEY. Never hard-code, log or commit the key. Never call the API from browser code: go through a server route or function.

## API
Auth header on every request: Authorization: Bearer $ZANI_API_KEY
Rate limit: 60 requests per minute per key, all endpoints combined. On HTTP 429 wait for the Retry-After header (60 seconds) and retry.

### POST https://www.tryzani.com/v1/text-to-speech
JSON body:
- text (string, required): 1 to 20000 characters. 1 character costs 1 credit, charged only when generation succeeds.
- voice (string, required): voice slug or id, e.g. "vjosa". List them with GET /v1/voices.
- dialect: standard | kosovo | gheg | tosk (default standard)
- emotion: neutral | warm | confident | energetic | calm | empathetic | serious | joyful | solemn (default warm)
- pace: slow | medium | medium-fast | fast (default medium)
- energy: low | medium | high (default medium)
- purpose: advertisement | audiobook | accessibility | social | documentary | meditation | education | news | podcast | narration | character (default narration)
- personality: free text, max 240 characters (optional)
- format: mp3 | wav (default mp3)

Response 200 (JSON): { id, status, queued, audioUrl, durationMs, characterCount, creditsCharged, format, voice: { id, slug, name }, error, createdAt, ... }
- status is one of "completed", "queued", "processing", "failed".
- Texts up to 900 characters finish inside the request: status "completed" and audioUrl set.
- Texts over 900 characters are queued: status "queued", audioUrl null. Poll GET /v1/generations/{id} every 2-3 seconds until status is "completed" (audioUrl set) or "failed" (error set). Stop after about 5 minutes.
- audioUrl is a temporary signed link (about 1 hour). Download the bytes and store them yourself if they are needed later.

### GET https://www.tryzani.com/v1/generations/{id}
Same JSON shape as above. Use it to poll queued generations. 404 if the id is not yours.

### GET https://www.tryzani.com/v1/voices
Returns { voices: [{ id, slug, name, gender, age, dialect, description, tags, badge, custom, sampleUrl }] }. sampleUrl is a ready-made audio sample of the voice.

### GET https://www.tryzani.com/v1/account
Returns { credits, plan, status }. Use it to check the remaining balance.

## Errors
Failures return JSON { "error": string } with these statuses:
- 400 invalid request (an "issues" array explains which field)
- 401 missing or invalid API key
- 402 not enough credits
- 404 voice or generation not found
- 429 rate limit exceeded (honour Retry-After)
- 500 / 502 temporary generation failure (no credits charged; retry once after a few seconds)
Error messages are written in Albanian: map them to your own user-facing messages instead of showing them raw.

## What to build
- A small wrapper module (for example lib/zani.ts or the equivalent for this stack) exposing speak({ text, voice, dialect?, emotion?, pace?, energy?, purpose?, format? }) that returns the audio bytes or a stored file path. It must handle polling, a 60 second request timeout, retry on 429 (Retry-After) and on 5xx once, and throw typed errors.
- Validate text length (max 20000) before calling. Split longer content on sentence boundaries into several requests.
- Cache audio for identical (text, voice, options) so credits are not spent twice.
- Add tests for the wrapper using a mocked fetch, plus a tiny command or script that generates one sentence so I can try it.
- Do not add new dependencies unless they are genuinely needed.

## Done when
An example call in this project turns Albanian text into an audio file, the key is read from ZANI_API_KEY, errors are handled, and you have told me exactly how to run it.