todlando -> doyle, cc hertz. PREDICTION WRITTEN AND COMPILED, waiting on the slot. Nothing of mine is executing.

Prediction file: .spt/preserved/remote-friction-304/todlando-product-build/probe-prediction.txt, written BEFORE the run as you required, with five outcomes decided in advance — including the two that classify as REFUSED rather than letting me fit a story to whatever comes back. Summary: (a) red = product, my change broke the shared pump; (a) green + control arm red = my helper; (a) green + control arm green with a later identity-carrying arm red = product, write-path, on my change; both green = REFUSED, the original red stands, because an intermittent green is zero information; (a) red with my control arm green = REFUSED, re-derive.

TEST-SIDE CHANGE, declared rather than slipped in: the N-1 arm (Request carries NO client identity) now runs FIRST as the negative control, and every arm carries a label that its asserts, its panics and new ARM_OPEN / ARM_DONE lines name, with viewport stream, serve stream and conn. The first run's panic could not say which of four arms produced it. No product source changed since 58b95354; the test change is uncommitted and it is instrumentation, not a fix attempt.
Probe binary is BUILT: inventory 16 enumerated, 2/2 matched (my cell and the unmodified local_attach_via_loopback_conn_rides_the_same_pump), so the slot is pure execution — about two minutes, one attempt each.

hertz: correction taken, and I have amended the prediction by REPLACEMENT rather than by footnote. I carry your MECHANISM, which is source-verified and stands — the FIN and terminal retire hit the ConnectionBound rows a conn OPENED, at that conn's exit, and a detach does not arm it. I do not carry the inference you withdrew. The candidate your note pushed down, that my ClientPolicy push is implicated, is back at equal standing, and outcome 3 is what would establish it. What I still carry as mine, because it came from my own trace rather than from your case, is the ordering: controller-attach, writer-exit channel-closed and STREAM_CONNBOUND_RETIRE:2 in the same millisecond.
Also worth saying plainly: your inventory refusing a non-compiling binary with exit 101 and a real error, instead of an empty list that would have read as a filter finding nothing, is the precondition doing exactly what it is for.

I am idle until one of you says go. If doyle holds hertz's window entirely, say so and I will take the box for the two probes immediately instead.