fix(speech): serialize pronunciation dictionary correctly

[kartikkabadi]

Summary

mmx speech synthesize --pronunciation built the wrong JSON shape for pronunciation_dict, so every request using the flag was rejected by the MiniMax T2A API (/v1/t2a_v2). Custom pronunciation — e.g. Chinese polyphonic characters (多音字) — was completely unusable from the CLI.

Fixes #187.

Before

"pronunciation_dict": [
  { "tone": "(chong2)", "text": "重" }
]

The CLI split each word/(pronunciation) value on / and rebuilt it as { tone, text } inside an array.

After

"pronunciation_dict": {
  "tone": ["处理/(chu3)(li3)", "危险/dangerous"]
}

Each value is preserved verbatim inside a tone string array on an object, matching the official PronunciationDict schema. Dry-run output matches the API-required pronunciation_dict shape. The payload shape is confirmed by the official API documentation and the reporter's raw API reproduction in #187.

Root cause

Two places encoded the wrong shape:

• src/types/api.ts — SpeechRequest.pronunciation_dict was typed as Array<{ tone: string; text: string }>, describing an array of objects.
• src/commands/speech/synthesize.ts — the --pronunciation block split each value on / and rebuilt it as { tone: to, text: from }, then wrapped the results in an array.

The API rejects the array form with Mismatch type open_platform.PronunciationDict with value array.

Changes

• src/types/api.ts — pronunciation_dict?: { tone: string[] } to match the API contract.
• src/commands/speech/synthesize.ts — pass each --pronunciation value through verbatim into body.pronunciation_dict = { tone: flags.pronunciation as string[] } instead of splitting it. No CLI-side validation of pronunciation syntax — the MiniMax API is the source of truth for pronunciation syntax.
• test/commands/speech/synthesize.test.ts — 2 dry-run regression tests (see below).

The change is narrowly scoped to issue #187 — no unrelated refactors.

SDK consumers

SpeechRequest is an exported type used by the public SpeechSDK. Any external code building the old Array<{ tone; text }> shape will get a compile-time type error after this change. That code was already broken at runtime (the API never accepted the array form), so this surfaces the break earlier — at compile time instead of at the API call. The SDK itself forwards SpeechRequest correctly via toMerged and requires no runtime changes. To migrate, pass { tone: ["word/(pronunciation)"] }.

Tests

Added 2 focused dry-run regression tests in test/commands/speech/synthesize.test.ts:

• --pronunciation serializes multiple values with the API-compatible shape — repeated --pronunciation values covering both the pinyin form (处理/(chu3)(li3)) and the plain-replacement form (危险/dangerous) shown in the official API docs. Asserts pronunciation_dict is an object with a tone string array (not the old array-of-objects shape) and that each value is preserved verbatim — no splitting, trimming, or rewriting.
• --pronunciation with a single value serializes under tone — single value (Omg/Oh my god, from the official API docs example).

Verification

• bun test test/commands/speech/synthesize.test.ts — 20/20 pass (2 new)
• bun run typecheck — pass
• bun run lint — no new errors (2 pre-existing errors in unrelated test files on upstream/main)
• bun test — no new failures (40 pre-existing failures on upstream/main are unrelated to speech synthesis; +2 tests, +2 pass)
• bun run build — pass

Reproduction

mmx speech synthesize \
  --text "我要去重庆出差。" \
  --pronunciation "重/(chong2)" \
  --out out.mp3

Before: Error: API error: invalid params, Mismatch type open_platform.PronunciationDict with value array
After: dry-run output matches the API-required pronunciation_dict shape.