After an agent-team session restart, audit artifacts instead of trusting the last status: a run's rc=0 can hide missing results

When a lead coding-agent session restarted, its background teammates no longer appeared in the agent list, and their last messages described work as queued or running. The status files were also stale. What gave the true state was checking artifacts directly: process lists on both machines, log and JSON timestamps, git status in each worktree, and CI runs. That audit surfaced things no status file reported. (1) A runner's exit line said rc=0 while its JSON held only an error (a JS exception in the benchmark script), so exit code alone was not success; read the result payload. (2) One run was killed by its own memory cap with almost nothing in its log, and the summary table still said pending, so a failure looked like work not yet done. (3) A half-finished merge from a side branch left many uncommitted files in the main worktree. (4) A check run with a deliberately broken guard printed PASSED; a mutant run that passes means the test does not catch that bug, so inspect mutant logs for the expected failure rather than for a green line. Useful habits: have runners write an explicit result status (ok, error, capped, slept) that a table is generated from, and on resume list processes, newest logs, dirty worktrees and CI before reporting anything.