CodeAbra/iai-mcp

Fixed

Idle detection survives invalid UTF-8 in system command output. All four subprocess readers in the idle detector (pmset -g log, ioreg, busctl, pmset -g) decoded stdout strictly, so a single non-UTF-8 byte — observed in a multi-megabyte power log — raised UnicodeDecodeError past the handler chain and cost the lifecycle tick its sleep-eligibility signal. They now decode leniently (errors="replace"), which keeps the check working: the sleep-marker scan and timestamp parse already tolerate replacement characters.

Reported with a full root-cause trace by @res-pstepan (#86), whose report also surfaced that all four readers shared the pattern, not just the one that failed.

The bench-harness smoke no longer turns red on a clean checkout. It armed itself whenever the Hugging Face caches and huggingface_hub happened to be present, then read a bench baseline JSON that is not tracked — so installing an unrelated package could flip the suite from green to a failure that looked like a bench regression. The module now sits behind an explicit --bench opt-in (same convention as --perf and --live), the baseline-drift check skips with a clear reason when the baseline has not been generated, and harness output goes to a temp directory instead of tests/fixtures/.