test_agent_engine_environment.py: 13 tests fail on macOS (GNU-only stat -c / base64 -d <file> in the local bash) #7778

🔴 Required Information

Describe the Bug:
13 tests in tests/unittests/integrations/vmaas/test_agent_engine_environment.py fail on macOS. The test's fake sandbox runs the commands that AgentEngineEnvironment sends to the (Linux) sandbox in a local bash. Since the chunked large-file read/write changes on 2026-10-27 (4c88040, 5843c58), those commands use GNU coreutils options that the BSD tools on macOS don't accept. CI runs only on ubuntu-latest, so this isn't caught there.

The production code is probably fine, because the real sandbox is Linux. This is a test-portability problem, but it means pytest tests/unittests is red for anyone developing ADK on a Mac.

Steps to Reproduce:

1. On macOS, pip install -e '.[test]' on main (571670d).
2. pytest tests/unittests/integrations/vmaas/test_agent_engine_environment.py

Expected Behavior:
The tests pass, or skip, on macOS.

Observed Behavior:
13 failed, for example:

FAILED tests/unittests/integrations/vmaas/test_agent_engine_environment.py::test_written_file_is_read_back
E       OSError: Failed to read .../test_written_file_is_read_back0/new dir/data.bin: stat: illegal option -- c
E       usage: stat [-FLnq] [-f format | -l | -r | -s | -x] [-t timefmt] [file ...]

Across the 13 failures the errors are stat: illegal option -- c (6) and base64: invalid argument <file> (7). The other GNU-only options in the same command strings are dd iflag=fullblock, mv -T and chmod --reference. Failing tests: test_written_file_is_read_back, test_empty_file_is_read, test_file_of_one_slice_is_read_in_one_command, test_large_binary_file_is_read_in_slices, test_symlink_to_large_file_is_read_in_slices, test_large_file_truncated_during_read_raises, test_write_is_split_at_chunk_boundaries[one_chunk_plus_one / exactly_two_chunks / two_chunks_plus_one], test_large_binary_file_round_trips, test_chunked_write_replaces_existing_file, test_chunked_write_keeps_existing_file_mode, test_chunked_write_with_long_multibyte_filename.

Environment Details:

• ADK Library Version: 2.13.0 (editable install of main @ 571670d)
• Desktop OS: macOS 26.6.2 (arm64), /bin/bash 3.2.57
• Python Version: 3.13.15

Model Information:

• Are you using LiteLLM: N/A (unit tests only)
• Which model is being used: N/A

🟡 Optional Information

Regression:
Yes. Before 5843c58 the only such command was printf %s ... | base64 -d > file, which works with macOS base64 because it reads stdin. The failing forms arrived with the large-file read/write support on 2026-10-27. On macOS, base64 -d <file> exits 64 with invalid argument, and stat -c %s <file> exits 1 with illegal option -- c.

Possible fixes (happy to send a PR for whichever you prefer):

1. Skip these tests when the local tools aren't GNU coreutils, e.g. a module-level pytest.mark.skipif that runs stat -c %s once and checks it succeeds. Smallest change.
2. Have the fake sandbox resolve GNU tools when they're installed (gstat, gbase64, gdd, gmv, gchmod from Homebrew coreutils) by prepending a shim directory to PATH, and skip only when they're missing. This keeps the tests running on Macs that have coreutils.

AI-assisted: yes (Cana)