fix: undici headers

[climba03003]

Fixes #287

Finally found the problem is caused by both http and undici.

1. http is not always treats the header to latin1.
   https://github.com/nodejs/node/blob/2a29df64645a70bbb833298423a29206c4ec6a2e/lib/_http_outgoing.js#L361-L373
2. undici always perform Buffer.from('').toString('utf8') regardless the header encoding.
   https://github.com/nodejs/undici/blob/2b260c997ad4efe4ed2064b264b4b546a59e7a67/lib/core/util.js#L216-L229

In combine of the above two problem.
The headers will parsed as a completely difference byte when chaining.

In this fix, we always treats the header as latin1.
So, we are always consistence in the encoding.

Checklist

• run npm run test and npm run benchmark
• tests and/or benchmarks are included
• documentation is changed or added
• commit message and code follows the Developer's Certification of Origin
  and the Code of conduct

[SimenB]

should there be an option in undici to leave headers as latin1 encoding?

EDIT: Ah, noticed nodejs/undici#1903 now

[climba03003]

I think yes. Otherwise, the patch is prone to error.

undici always trans-code the header. Maybe it should really detect the encoding before trans-code. That means the behavior should be identical to http.request.

Actually, I already open a issue on undici.
See nodejs/undici#1903

[mcollina]

lgtm

[mcollina]

Would this need a revert after undici is fixed?

[climba03003]

I believe, yes.
We can update the patch function to detect undici version once it fixed.
It can maintain the best compatibility.

[arturkin]

FYI this causes regression with set-cookie headers in case it's an array of values. Had to revert to the previous version due to this.

[climba03003]

FYI this causes regression with set-cookie headers in case it's an array of values. Had to revert to the previous version due to this.

Can you provide a reproducible code in a separate issue?

[SimenB]

The change can probably be reverted together with a bump of the undici dependency?

[climba03003]

The change can probably be reverted together with a bump of the undici dependency?

Even with bumping undici in dependencies, I believe it wouldn't catch-up the latest.
I am fine if it would also update undici when fetching new package.

[SimenB]

It would

[mcollina]

Could you open a fresh PR?