Can the agent distinguish stdout and stderr?

Hi!

I find that some test functions failed against a generated executable because the test expected to have some string in **stderr**, while the executable gave that string in **stdout**.

I also noticed that ProgramBench’s container execution helper appears to merge stdout/stderr streams [here](https://github.com/facebookresearch/ProgramBench/blob/main/src/programbench/container.py#L85).

Is stdout/stderr separation intentionally unavailable to agents during exploration? (I think it is still possible for an agent to manually test stream placement with shell redirection, e.g. `cmd >/tmp/out 2>/tmp/err`, but the default observation format seems to make the distinction easy to miss.)

Thank you!

### The commands I ran
```bash
uv run --with mini-swe-agent mini-extra programbench \
--filter "abishekvashok__cmatrix.5c082c6" \
--output output \
--model openai/gpt-5.4

uv run programbench eval output
```

Score log
```bash
Evaluation Summary
Instance Score Comment
abishekvashok__cmatrix.5c082c6 94 507 tests
Average 94 1 instances
```

### A failed test in eval.json

- task: `abishekvashok/cmatrix`
- test function name: `eval.tests.test_cmatrix.TestColorOptions.test_color_invalid_shows_error_message`
- branch: `1b991a57d4e9`

```python
self =

def test_color_invalid_shows_error_message(self):
"""Test invalid color name produces error message."""
result = run("-C", "purple")
assert result.returncode == 0 # Exits with 0 but shows error
> assert b"Invalid color" in result.stderr
E AssertionError: assert b'Invalid color' in b''
E + where b'' = CompletedProcess(args=['./executable', '-C', 'purple'], returncode=0, stdout=b' Invalid color selection\n Valid colors are green, red, blue, white, yellow, cyan, magenta and black.\n', stderr=b'').stderr

eval/tests/test_cmatrix.py:120: AssertionError
```

Inizia in src/programbench/container.py intorno alla riga 85, dove l’helper per l’esecuzione del container sembra unire stdout e stderr. Riproduci il comportamento con il caso di valutazione cmatrix in eval/tests/test_cmatrix.py:120 e confronta gli stream osservati con l’output effettivo del comando. Il lavoro è completo quando il comportamento previsto di stdout/stderr è chiaro e l’osservazione predefinita dell’agente non oscura più la distinzione, dopo aver verificato il comportamento di valutazione rilevante.

Scritto dal modello di indicizzazione a partire dal testo della issue.