fix(cli): don't truncate --format json output through a pipe

[DanMat]

Closes #279 (a recurrence of #184)

--format json output is truncated when stdout is a pipe and the exit code is non-zero. Two root causes, both fixed here:

1. write() resolved on the source Readable's end event, not on stdout draining — out.write()'s return value was ignored, so it resolved with bytes still buffered. It now uses out.write(data, callback) and resolves when the chunk is actually handled, delivering the flush its comment promises.
2. The JSON branch then called process.exit(exitCode), terminating before a buffered pipe finishes flushing. It now sets process.exitCode and returns, matching the human-readable branch (which already worked for exactly this reason).

Added a regression test for write() that fails on the old implementation and passes on the new one. Changeset included (patch).

Verified: full cli suite (31 tests, incl. the existing -f json snapshot) passes; prettier + tsgo clean.

Written with AI assistance; I've reviewed and tested the change and will maintain it.

[changeset-bot]

🦋 Changeset detected

Latest commit: 5d03a30

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@arethetypeswrong/cli	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[DanMat]

Reproduced the truncation locally and confirmed the fix. Fixture: a failing package with 17 entrypoints (≈311 KB of --format json output), piped through cat:

Before (base main):

$ attw pkg.tgz --profile esm-only --format json | cat | wc -c
131072        # stops at a pipe-buffer boundary — JSON.parse() fails (truncated)

After (this PR):

$ attw pkg.tgz --profile esm-only --format json | cat | wc -c
311241        # complete — JSON.parse() succeeds

Writing to a file (> out.json) yields the full 311,241 bytes on both, so the truncation is governed by the pipe + non-zero exit, exactly as described in #279. The 131072 here is 2×64 KiB; the exact boundary varies by platform, the mechanism doesn't.

The added unit test also fails on the base write() implementation and passes with the fix.