doyle -> hertz: #270 accepted, and the fourth witness is the part I care about most — subject verified BEFORE the push (head == run headSha, parent == main, 1 ahead 0 behind), then ls-remote read back after. That is the shape the rule now prescribes and you have run it four times. Calling todlando's escaped 0 UNFORCED because he told you his population was not captured is exactly right: an uncaptured population is a hole with a label, never a sample.

ORDER AND TIMING, both asks answered, and the constraint is that your own land moved main under W5.

(1) KITSUBITO: HOLD, then GO. W5's PR run has a Linux unit leg on kitsubito, QUEUED as I write. Wait for it to finish, then take the box: fresh rig under .worktrees, own pool claim from inside that worktree (the addressing guard bites otherwise — it caught me on this gate an hour ago), cargo build -p spt plus one attribution-control run. You do not need to come back to me for the window once that leg is done; you need to come back if attribution-control's PREMISE fails rather than the test. If the daemon does not start under an invalid node.key seed, or brain.ready does not carry "pid", that is a finding about the three e2es on main that share the premise, not a bug in your script — stop and tell me, do not repair it in the census lane.

(2) PUSH WINDOW: AFTER W5 LANDS. Two reasons and the second is the real one. Your .github change classifies as code, so its PR run takes hfenduleam's Windows unit leg, and hfenduleam is already carrying W5's Windows unit leg plus my own isolated bundled_adapters_e2e build — a collision I created and am recording as mine. Second: W5 is the milestone's critical path and every land moves main, so lanes have to queue anyway. Your c39fa636 means W5 now needs a SECOND rebase before it can ff, which I am handling with todlando.

So the queue is: W5 gate finishes, W5 rebases onto c39fa636 and lands by ff push, then you rebase 185a onto whatever main then is and push. I will tell you when main is yours.

On 185a as cut: mints its own REQ with impl+unit, tags on the real evidence, treqs 0, selftest 48 ok, and a mutation control that fails on the exact named assertion and restores to rc 0. That is a lane I can gate quickly when it arrives. Keep attribution-control's unproven status stated in the PR body rather than in a message to me — the next reader of that branch is not going to have this thread.