[next/nightly] Incorrect _'obj.prop' is possibly 'undefined'._ Error in CFA with `exactOptionalPropertyTypes` and Optional-`never` Properties

[SlurpTheo]

🔎 Search Terms

exactOptionalPropertyTypes

🕗 Version & Regression Information

• This changed between versions typescript@5.2 and typescript@5.3.0-dev.20230816 (still happening in typescript@5.3.0-dev.20230825)

⏯ Playground Link

https://www.typescriptlang.org/play?exactOptionalPropertyTypes=true&ts=5.3.0-dev.20230825#code/KYDwDg9gTgLgBAMwK4DsDGMCWEVwLYCGmKAFAQFxwDOMUxA5gNoC6cAPnKgCbALHBcAlHADeAWABQcaXDQ4acAF5wAvHAJwA9Jsky9cAPyj1cAL6790yiLgAjSowBEMABbAowR61MBuSRelFADoNbQD9I2CCIJgIAGVaBhJBcL1KYNsY+MSUemSfLU04GCoAAQBWIIBmOHcoaDgcOAADDOa4Nw84AHc3XGbQAgwAeTAsHAIAGwAFerB3GABPABVF+ap2zCpalAJbSYF-KRltHugAawBaAnruOABiAEZwuRQFRQB1C4BBW5QuazqSg0Oi5FgAGjsBko3F4-C4ZnYxns1ByTGYkII0M4-zhKAEiLUij8x0CXyg51+EG4IUKqRkkXJlL+XBCWQSoLyKVJ+nSTKpNMysQ5SUEJPCp26F2uLIeACZwo4CI44MQlIVii4tqrtgAqXXDAByABkAJr6kwANymmARUopN2p-x6fRagxGY2wuxmcwWKzWwA2cH1AEk4hbgLt9oceXpImzhWjkvSrEogkLspz8pJzBIgA

💻 Code

export function main(a: string[] | undefined) {
    const z = a //
        ? { a }
        : { b: ["there"] };

    z.a //
        ? z.a.toString()
        : z.b.toString(); // ts@5.3 error on `z.b` here when `exactOptionalPropertyTypes` is enabled

    // work-around #1
    const zWorkAround: { a: string[], b?: undefined } | { b: string[], a?: undefined } = z;
    zWorkAround.a //
        ? zWorkAround.a.toString()
        : zWorkAround.b.toString();

    // work-around #2
    "a" in z // this is **ONLY** a valid workaround when `exactOptionalPropertyTypes` **IS** enabled
        ? z.a.toString()
        : z.b.toString();
}

🙁 Actual behavior

Errors in code

'z.b' is possibly 'undefined'.

🙂 Expected behavior

Try any pervious version (I checked back to typescript@4.0) and the CFA from the !!z.a branch works fine.

Additional information about the issue

An interesting little problem/tidbit back to typescript@4.4 is this IntelliSense:

• 

Why is a optional in the "a" in zWorkaround2 branch?

[Andarist]

It might be the regression from #54777 that im investigating