Increase the default auto alllocate chunk size

[jasnell]

From 4KB to 64KB. This increases the default chunk size for default reads on internal streams, which should reduce the number of individual reads for larger payloads when piping data or iterating over the stream.

[ask-bonk]

I'm Bonk, and I've done a quick review of your PR.

This PR increases the default auto-allocate chunk size for byte streams from 4KB to 64KB to reduce read count for larger payloads.

1. [HIGH] Ungated behavioral change: The default chunk size for byte-stream default reads is user-observable (it determines the buffer size passed to pull() via byobRequest.view and the chunk sizes returned by reader.read()). Changing from 4KB to 64KB without a compat flag is a breaking change for deployed workers that depend on the old size. The existing noAutoAllocateChunkSize flag controls whether the default is applied, not what value it is. This needs its own compat flag or should be gated behind an existing one.
2. [LOW] Stale comment: The legacy-path comment still says "default to 4096".

---

This review was generated by an AI assistant and may contain inaccuracies.

[ask-bonk]

I'm Bonk, and I've done a quick review of your PR.

This PR increases the default auto-allocate chunk size for byte streams from 4KB to 64KB. I posted a review on PR #6534 with the following findings:

1. [HIGH] Ungated behavioral change: The auto-allocate chunk size is observable by user code — it determines the buffer size in byobRequest.view during pull() and the chunk sizes returned by reader.read(). Changing this 16x without a compatibility flag affects all existing workers on the legacy code path that don't explicitly set autoAllocateChunkSize. The existing noAutoAllocateChunkSize compat flag only controls whether a default is applied, not what value it is. This needs a new compat flag.

2. [LOW] Stale comment: The comment at standard.c++:3041 still says "default to 4096" and the comment in common.h:118 now reads 64 * 1024 where "64KB" would be more readable.

github run

[npaun]

Do you have data from local testing, or are we going to have to try it live to see if helps?

[jasnell]

Up to 80% faster TTLB on a 1 MB payload.