#299 ATTRIBUTION RULED from your run + perri's numbers. No second run. Slow arm = (a): 10105/10120 ms is PRESENCE_PROBE_CEILING (10000) + ~110 ms of everything else; your 457 ms pre-subnet block is bounded and cannot reach 10 s; perri's 3822 is a dial resolving late. One unlisted peer, one probe, two outcomes: your dial errored at 628 ms, perri's hung to the ceiling. Network deadline — not contention (registry 68 ms, none measured), not metadata (457+68 ms, sub-second).
457 ms pre-subnet: NOT chased in this lane. Record it in your #299 closing comment as the largest unattributed metadata cost in a fast run, bounded, not the complaint. Budget disagreement (dial carrier 30 s vs ceiling 10 s, detached dial outlives the cutoff): same — record it, no fix here; it does not touch wall time.
REMEDY: the contract (REQ-UNLISTED-PRESENCE-PROBE, CONTEXT.md presence-probe entry) promises a BLOCKING active probe, so full removal of the wait is a contract change and is with the operator now. Two shapes I put to them: (A) keep blocking, ceiling parameter 10000 -> 2500 ms = the binary's existing PROBE_TIMEOUT value (the parameter stays separate; the const stays untouched) — a probe past 2.5 s reads UNKNOWN never OFFLINE, exactly as today past 10 s; REQ title + CONTEXT.md sentence amended to the new number with this measurement as the justification; the existing 'wedged probe costs ONE ceiling' cell re-pinned. (B) contract change: list renders without waiting, presence word 'not asked' unless --probe. I recommended A. Do NOT start either until I relay the pick. Keep #49/#267 (finish the CLOSE_WRITER_GONE rename before check-green counts) and then #300 moving meanwhile — your lane does not wait on this.