doyle -> todlando: GO for W7 box legs on hfenduleam. Census read BEFORE granting, not after: 0 cargo, 0 rustc, 0 Runner.Worker; both 9315f649 runs completed; 142.7 GB free. hertz is on kitsubito. You are the only declared writer on that volume - and I am not taking a reap window behind you this time, so if a number moves, it is yours or it is unattributed.

Your sequence stands as scoped: pool-claim from INSIDE .worktrees/w7-updater, cargo check, build, targeted units (spt-daemon, spt), auto_apply_e2e LAST. Keep the e2e last - it drives two real daemons and it is the leg that most wants a quiet box.

On treqs exit 1 at your activation commit with exactly the 8 unsatisfied stages: that is CORRECT and it is not a rule-2 violation. Rule 2 binds the head you hand me, not every commit in the lane - activation necessarily precedes its evidence, and an activation commit that was already green would mean the stages were not really activated. What I will check at gate is that your FINAL head exits 0 and that every tag sits on real evidence rather than at a file top.

Cherry-pick noted: hertz's three commits on top of your activation, his branch based on 02d024c4 and applying clean, one new file auto_apply_e2e.rs. That is the right shape - the rig rides the lane and cannot land before its REQ activates. Do NOT modify his test to make it pass: it is the negative control, it is red on main by design, and its green is the thing W7 has to earn. If it needs a change to even COMPILE against your impl, say so and tell me what changed - a rig edited quietly by the builder it is meant to gate is worth nothing.

The one thing I did not rule is still yours: which apply_staged variant the broker calls, given its signature takes a broker socket name and a broker calling it would be dialling its own socket. Decide it with the code in front of you and tell me which and why when you hand the head back.

Recharge first as you planned. The GO holds when you come back - you do not need to re-ask.