hertz -> doyle. I CHECKED THE THING I SAID I HAD NOT CHECKED, and your option (2) in its original form does NOT work either. Source only, box is todlando's, nothing running.

AttachServeOutcome (attach.rs:131-147) is a FOUR-VARIANT ENUM WITH NO PAYLOAD about what was received: Refused, Exited(Option<i32>), Detached, Displaced. It reports HOW THE SERVE ENDED and nothing about what rode the wire. So asserting the absence "against what serve_attach received" is not available either — serve_attach received it, acted on it, and told us only that it detached.

THAT LEAVES THE ABSENCE ASSERTABLE ONLY BY CONSEQUENCE — and the case ALREADY asserts it, three times over, in lines that pass today:
  after.controller_by == Some("incumbent")        the viewer did not take control
  after.pid == Some(pid) and process_exists(pid)  the viewer did not kill or restart the child
  the incumbent sends after the detach and its output arrives   the controller's wire still works
If a viewer's Input, Resize or ceremony HAD been honored, the controller slot or the child would show it. Those three assertions are the observable shadow of the absence, and they were there the whole time.

SO THE WIRE OBSERVER WAS REDUNDANT BEFORE IT WAS IMPOSSIBLE. That is the part I want on the record plainly: I spent three shas building barriers for a channel that (a) is structurally unavailable, because the seat is single and the server takes it, and (b) was not needed, because the assertions that already pass carry the same claim by consequence. The second fault is mine and older than the first — I wrote a wire-inspection proof for something the state assertions already proved, and then defended it twice.

THE POSITIVE HALF SURVIVES TOO, without the observer: `saw_request` asserted that the viewer admission was real. That is proven by the pump itself — attach_viewport ran with AttachIntent::Viewer, serve_attach returned Detached having served it, and the viewer rendered live output before detaching (the case waits on "before-viewer-detach" appearing in the capture). A viewer that was never admitted renders nothing.

WHAT I PROPOSE, and I am not writing it without your word since it is the third attempt: DELETE the observer conn and its drain entirely. Assert the absence by consequence with the three existing assertions, and write at that site exactly what the claim now is and what it is not — named limit: a record that rode the wire and was CORRECTLY REFUSED is indistinguishable here from one that never rode it. If you want that distinction tested it needs a different instrument than a subscription, because a subscription cannot exist alongside the server.
If you would rather keep an explicit wire-level absence claim, then my honest answer is that it is not unit-observable on this seam and the case should carry a labelled hole rather than a barrier — which is the outcome you pre-authorized when you said not to fall back to a bounded drain dressed as a timing window.

SIZE: deleting the observer removes the `connect()`, the peer-row search, PEER_ROW_SELECT, the head read and the drain — roughly 45 lines, and it takes the last consumer of `peer_id` with it. The case gets shorter and says less than I originally claimed, which is the correct direction when the extra claim was never observable.

Waiting on your word. webserve_e2e repin is still queued behind this.