client: return one-frame response bodies without copying

[allada]

collect_body_bounded copied every frame into a BytesMut, including the common case of a body that arrives whole in one frame. Hold the first frame by reference count and return it directly when no second frame follows; promote to a buffer only when one does.

Using http_body_util::Limited here instead would need B::Error to be std::error::Error rather than Display, which propagates to ServerStream and BidiStream and the transport trait's body. Not worth a public API break for this.

[iainmcgin]

[claude code] Thanks — the rewrite is correct on the points that matter: the running len equals the old buf.len() at every check (empty frames add zero), the > max_size boundary is unchanged so a body of exactly max_size still succeeds, nothing is retained before the limit passes, and trailer frames behave as before. Full hands out its stored Bytes unchanged, so the pointer-equality test is valid and can't false-pass while src is alive.

Two things before it can merge:

1. It no longer compiles after a rebase onto main. main gave collect_body_bounded a third deadline: Option<Instant> argument (post-deadline body errors are classified as deadline_exceeded, #238), and the merge is textually clean because the hunks don't touch that arm — but the two new tests call the two-argument form and fail with E0061. , None on both fixes it locally; the rest of the gate then passes. Please also keep the classify_body_read_error arm intact when you resolve, since taking the PR side of that region would silently revert it.

2. Say that the returned Bytes may alias the transport buffer. For a one-frame body the value handed back is hyper's frame slice — a piece of the h2 codec buffer, or of the h1 adaptive read buffer, which grows to several hundred KiB — and the Connect unary path keeps it for the lifetime of the decoded response. That matches what http_body_util::Collected::to_bytes does, and we're going with zero-copy by default rather than a size threshold or a config option; the caller-local remedy is to copy what you retain. So it needs to be in the doc comment on collect_body_bounded, and the changelog should say it rather than naming collect_body_bounded itself, which is private and not something a user can look up. (The changelog's "exactly-sized buffer" is also only exact for a two-frame body; a third frame reallocates. "Sized for the frames seen so far", or drop the clause.)

Smaller:

• The data.is_empty() { continue } branch has no test. With ChannelBody, frames ["", "foo"] should return "foo" by refcount, and ["foo", ""] should not promote to a BytesMut.
• The multi-frame test's non-aliasing assertion compares against a from_static pointer in rodata, which a heap buffer can never equal, so it doesn't catch anything the b"foobarbaz" equality doesn't already; fine to drop.
• The final match relies on head being None whenever joined is Some (true, via head.take()); a debug_assert! pins it.
• "nothing else pins it" in the test doc isn't quite right — src holds the second refcount, which is what makes the assertion safe.
• Fragment timestamp is hand-typed; task changelog-new generates one.

This stacks cleanly on #264. Iain will do the final review and merge.

[iainmcgin]

[claude code] Pushed 4711503 onto your branch with the changes from the review, so this is ready to merge once CI passes:

• Merged main. The two new tests now pass None for the deadline argument (E0061), and the classify_body_read_error arm is intact.
• The doc comment on collect_body_bounded now says a one-frame body may be a slice of the transport's read buffer.
• Added tests for an empty frame before and after the only data frame, and for a multi-frame body being joined into a new buffer. Dropped the from_static pointer assertion and added a debug_assert! that head and joined are never both set.
• Regenerated the changelog fragment (Changed-20260924-190142.yaml). It is scoped to Connect unary and client-streaming responses, names OwnedView, says to_owned_message() still aliases bytes fields, and links this PR.

Thanks for the change.