test_bench_exits_nonzero invokes full bench CLI; hangs locally when Ollama is running

[adris-misra]

Description

The test test_bench_exits_nonzero invokes the full industrial-agents bench CLI to verify non-zero exit on failure. Since IA-4 and IA-6 now
make real LLM calls, this test runs the entire benchmark pipeline when
Ollama is reachable locally — taking minutes or appearing to hang.

In CI (no Ollama), the bench fails fast with a connection error → exit
code 1 → the test's assertion passes. So CI is unaffected, but local
pytest tests/unit -q is slow or hangs.

Surfaced during PR #15 (IA-4/IA-6 implementation).

Steps to reproduce

1. Ensure Ollama is running locally (e.g. llama3.2:1b pulled and serving)
2. From the repo root: pytest tests/unit -q
3. Observe the run stall at test_bench_exits_nonzero while it executes
   the full IA-4 + IA-6 benchmark against the live model

Expected behaviour

A unit test should never invoke the real benchmark pipeline or make live
LLM calls. The test should mock the CLI's bench invocation so it verifies
the non-zero exit path deterministically, in milliseconds, regardless of
whether Ollama is running.

Actual behaviour

The test invokes the real industrial-agents bench command. With Ollama
running, this executes IA-4 and IA-6 (real LLM calls), taking minutes or
hanging. Without Ollama (CI), it fails fast with a connection error, so
the assertion coincidentally passes — masking the design flaw.

Suggested fix: mock the bench invocation in this test, OR mark it
@pytest.mark.integration and exclude it from the default unit run.

Framework version

v0.1.0-pre (bench/iabench-synthesis-cost @ 66bc8fe)

LLM provider

ollama

Environment

Windows 11, Python 3.12