[akynte]

TestSIGKILLResume now writes down each step it checks and compares the sequence with internal/orchestrator/testdata/crash-replay.json. Any difference fails the test, so a passing CI run proves the file is what a real kill and resume produced.

Steps recorded (each is a value the test reads, not a description):

• the task's state after creation;
• runner A, in its own process, has written a file and is mid-turn;
• SIGKILL, and what Wait returns;
• the state the dead runner left: lease owner, phase, attempt;
• runner B's refusal while the lease is fresh;
• the stale-lease timeout (50 ms in the test, 2 m in production);
• what runner C's agent is told on resume, and whether the killed attempt's file survived;
• the reconciled attempt, the task's event log, and the final state.

Task ids and wall-clock times are normalised, so the file is identical on every run (checked 10 times locally, and under -race). -update-replay rewrites it when behaviour changes on purpose.

Test-only change: no production code is touched. go vet ./... and go test -race ./... pass.