[M] Accept any BufferSource in TextDecoder.decode

[evanrichards]

The problem

TextDecoder.decode returned an empty string for an ArrayBuffer and for a
DataView. The polyfill read input.length. Neither type has a length
property, so the loop never ran. The method did not throw, so a caller lost
the data without a signal.

The backend runs customer-supplied and agent-authored TypeScript in this
sandbox. A script that calls new TextDecoder().decode(someArrayBuffer) gets
an empty string today, and the failure appears far away from the cause.

const bytes = new TextEncoder().encode('hello world')

new TextDecoder().decode(bytes.buffer)            // '' before, 'hello world' now
new TextDecoder().decode(new DataView(bytes.buffer)) // '' before, 'hello world' now
new TextDecoder().decode(bytes)                   // 'hello world' before and now

The change

The WHATWG encoding standard defines the input as a BufferSource, which is
an ArrayBuffer or a view on an ArrayBuffer. The new function
toUint8Array converts the input first. The conversion keeps the byte offset
and the byte length of a view, so new Uint8Array(buffer, 6) still decodes
the correct bytes. An input that is not a BufferSource now throws a
TypeError instead of a silent empty string.

The method also builds the intermediate string in chunks of 8192 bytes with
String.fromCharCode.apply. The previous code called String.fromCharCode
one time for each byte. The chunked loop keeps the same escape and
decodeURIComponent pair, so the output and the error behaviour do not
change.

The measurement

A benchmark in the sandbox decoded 600 KB, three runs, in milliseconds:

decoder	ASCII	unicode
one call for each byte (before)	131-135	146-207
chunked calls (this change)	19-26	20-22
a hand-written UTF-8 decoder	75-101	77-80

A hand-written UTF-8 decoder removes the Annex B function escape, but it is
slower and it changes the behaviour for invalid UTF-8. This change keeps the
escape pair.

The tests

src/test/sync/util.test.ts and src/test/async/util.test.ts each get 12
cases: an ArrayBuffer, a DataView, a Uint8Array, a view with a byte
offset, a view with a shorter byte length, a signed Int8Array view,
multibyte characters, an empty input, a payload larger than one chunk, and
the two TypeError cases. Five of them failed before the change.

The version

package.json moves to 3.2.1. The fork needs a v3.2.1 GitHub Release after
this merges, so that @loop-payments/ts-sandbox can take the fix.

🤖 Generated with Claude Code

[evanrichards]

This change is part of the following stack:

• [M] Accept any BufferSource in TextDecoder.decode #3 ◀

_{Change managed by git-spice.}