- Parse the drained subprocess tail so the verdict is read when all output
lands in one final read (single-turn answers were silently dropped)
- Give each parallel worker its own .eval-runs/<id>/.claude/commands/ dir
with a uuid-unique command file so N runs no longer compete for one slot
- Never treat a non-skill tool call or block completion as a verdict; only
the final result event or drained tail is conclusive
- Add 10 tests covering all three bug classes (8 failed on the old code)
2a9862674d