jobs: only settle may close; e2e proof harness

close_all no longer joins pump threads from the owner thread: job:close()
joins the pump, and a pump parked in an owner-posted tool batch needs the
owner back in the loop, so a synchronous close at turn_end could deadlock.
close_all now cancels everything, closes only already-settled jobs, and
defers the rest to their own settle; jobs.active() ignores closing children
so the next turn can resume the id.

e2e/ reconstructs the end-to-end proof against the real binary in print
mode: a scripted fixture protocol plays primary and children, and the suite
asserts the product surface from process output and JSONL bytes on disk —
parallel batch with unknown-agent error, lane identity, tool inheritance
minus subagents.*, resume with manifest re-identification, workflow diamond,
sandbox fan-out with an ephemeral structured worker, catalog round-trip,
child-store isolation from the session picker. mise run e2e; check stays
independent.

1 files changed, 5 insertions, 1 deletions

diff --git a/spec/test_init.lua b/spec/test_init.lua
index 800dcaa..338b7a7 100644
--- a/spec/test_init.lua
+++ b/spec/test_init.lua
@@ -115,7 +115,11 @@ return {
                 handle.emit("turn_interrupt", { phase = "interrupt" })
                 assert(job._cancel_requested, "an interrupted turn asks its children to stop")
                 handle.emit("turn_end", { phase = "end", reason = "interrupted" })
-                assert(job._closed, "the end of the turn joins every child")
+                assert(not job._closed, "the end of the turn never joins a pump from the owner thread")
+                -- The pump exits (the fake settles cancelled on its next poll)
+                -- and the drain that notices it does the close.
+                jobs.await({ started }, "all")
+                assert(job._closed, "a child closes as soon as its pump exits")
             end)
             jobs.close_all()
             assert(checked, failure)