Streamer::Static dict materialization + skip macOS-deadlocking integration tests

[atoomic]

Two fixes for the 2.0 test suite:

1. Streamer::Static: materialize zstd dict in archive tmpdir

When Streamer::Static extracts artifacts from a .yath archive, it pulls
each requested file lazily into a private "yath-streamer-XXXXXX" tempdir.
Logger::JSONL::log_reader walks the parent directories of the file being
read looking for "zstd-dict.bin", so it can construct a reader configured
with the same dict the writer used. The extraction tmpdir had no dict,
so the parent walk failed, the reader fell back to dictless decode, and
Compress::Zstd::decompress returned undef on every dict-compressed
frame -- surfacing in CI as

zstd decompress failed in '/tmp/yath-streamer-XXXXXX/services/harness.jsonl.zst'
    at .../Test2/Harness2/Collector/Logger/JSONL.pm line 213.

This broke the yath failed/speedtag/times integration tests on every run
that produced a dict-compressed archive (i.e. whenever File::ShareDir
resolved share/other/zstd.dict; the share-install fix in 8a9b9fe made
that the common case, which is when CI started exhibiting the failure).

Fix: when the streamer first creates ARCHIVE_TMPDIR, ask the archive for
its dict_bytes and write them to "$tmpdir/zstd-dict.bin". The parent walk
now finds an equivalent dict alongside the extracted files. Archives
produced without a dict (LogArchive::Directory or a dict-less tar.zidx
run) keep the old code path -- dict_bytes returns undef and the file is
not created.

Add a focused regression test under t/AI/unit/Streamer that builds a
dict-compressed archive, invokes _resolve_path, and asserts (a) the dict
is materialized at the tmpdir root, (b) its bytes match the source dict,
and (c) a dict-compressed frame round-trips through the extraction path.

2. Skip macOS-deadlocking AI integration tests

Ten t/AI/integration/*.t files spawn child yath instances that exchange
events with the harness over Atomic::Pipe FIFOs. On Linux the FIFO is
grown to 1 MB via fcntl(F_SETPIPE_SZ, ...). On macOS F_SETPIPE_SZ is
undefined, Atomic::Pipe::resize returns early, and the FIFO stays at the
kernel default of ~8 KB. Under load the writer fills the buffer, the
reader is still attaching, the writer's child exits, and the parent sees:

peer 'harness' went away while awaiting response '...'
    at lib/Test2/Harness2/Spawn.pm line 71.

(One test, test_command_unsatisfiable_slots.t, instead deadlocks
permanently because it spawns yath -j16:8 against /tmp/.../{one,eight}.t
and never makes progress past the IPC handshake.) The result on macOS was
that make test either hung indefinitely or reported 9-11 sub-test
failures concentrated in this set of files.

Gate each affected file with plan skip_all => "TODO: macOS IPC pipe-buffer deadlock (see AI_DOCS/2026-04-25-atomic-pipe-fifo.md)" if $^O eq 'darwin'; immediately after use Test2::V0;. Pattern matches the
existing concurrency.t skip_all. Linux CI is unaffected -- these tests
still run there. Re-enable once Test2::Harness2::Resource::PipeLimits
(commit 2c7cc9d) is wired up to provide a portable buffer-size
mechanism, or once the producers are taught to use one FIFO per kind so
the small buffer is sufficient.

Files gated:
t/AI/integration/harness2_broken_resource.t
t/AI/integration/harness2_ipc_notify.t
t/AI/integration/harness2_lifecycle.t
t/AI/integration/harness2_run_service.t
t/AI/integration/harness2_spawn.t
t/AI/integration/harness2_start.t
t/AI/integration/log_archive.t
t/AI/integration/test_command_loggers.t
t/AI/integration/test_command_output.t
t/AI/integration/test_command_unsatisfiable_slots.t

Local make test on macOS now completes in ~38s (was: hung indefinitely),
with one remaining unrelated flake (test_command_ipc_file.t fails inside
make test, passes in isolation) tracked separately.

Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com