Fix ARG_MAX overflow in stderr blob test on Linux CI

[yasyf]

Summary

tests/test_proc.py::TestArunCli::test_stderr_newline_free_blob_past_64k built its child command by interpolating a 200,000-byte literal into argv through an f-string repr. On the Linux CI runners, argv plus the runner's large inherited environment exceeds the kernel's ARG_MAX, so execve fails before the child starts:

tests/test_proc.py:132: in test_stderr_newline_free_blob_past_64k
    await arun_cli(argv, stderr_tee=seen.append)
spawnllm/proc.py:226: in arun_cli
    proc = await asyncio.create_subprocess_exec(
E   OSError: [Errno 7] Argument list too long: '/home/runner/work/spawnllm/spawnllm/.venv/bin/python'
1 failed, 420 passed, 2 skipped in 8.07s

The child now generates the blob itself (sys.stderr.write('e' * 200_000)) instead of receiving it, so argv stays small. The test's subject is unchanged: it still asserts that a newline-free 200,000-byte stderr blob past the 64 KiB readline limit is captured whole rather than lost to a LimitOverrunError mid-drain, and still checks the exit code, exc.value.stderr, and the teed bytes.

This is a pre-existing break on main, not a regression from any open branch. The test arrived with #6 ("Fix stderr drain crashing on newline-free blobs past 64 KiB"), and every ci.yml run on main since has failed, blocking CI for every PR in the repo. It is not a flake: it fails deterministically on both the 3.13 and 3.14 jobs, so a rerun does not clear it. It passes on macOS, where argv has more headroom against a smaller inherited environment, which is why it went unnoticed locally.

Test plan

• uv run --extra dev pytest tests/test_proc.py -k stderr passes locally.
• The Linux CI run on this branch is the real proof, since the failure is environment-dependent.

https://claude.ai/code/session_014gKfZvjihGgFrEVCr3NJJb