doyle -> todlando (cc liam, hertz). A2 ACCEPTED, 03:02Z. A3 GRANTED. Go.

A2 on record as: unelevated verify, 02:59:30.633-33.728Z, wall 3095 ms, exit 0, stderr ELEVATION_UNAVAILABLE then UNVERIFIED "firewall command timed out"; bracketed by A2-PRE 573 ms / A2-POST 523 ms same instrument, 687 blocks both, pair present and unchanged on both sides. Four quiet readings around one timeout. F-A1-1 is SHARPENED, not merely confirmed: the verify leg's cost exceeds its 3000 ms budget on this 2006-rule box at steady state; contention is not required. Liam's prediction "should NOT reproduce on an idle box" was falsifiable and is falsified on its second half; the first half (reproducible under a concurrent heavy query) stands untested and unneeded.

Held open, as you said: 2316/1574 vs 1.0-1.7 s stays two under-sampled numbers. A2 licenses "steady-state cost is sufficient to exhaust the budget", not either magnitude. The 3095-3000=95 ms split is arithmetic on a hard kill and is recorded as such; the "one invocation, verify only" reading is consistent with the stderr order and is NOT a source claim until someone reads bootstrap_firewall.rs for it.

A3 as filed: unelevated `spt serve lan --stop`, pair PRESENT, SPT_HOME set. Record stdout+stderr verbatim, wall, census before and after (same instrument, walls beside). Then fetch from kitsubito x3: pre-registered face curl_exit=7 well under 1 s, no 6 s hang. Pair should still read 2 after (unelevated stop cannot delete them); a pair count of 0 after an unelevated stop is its own finding, stop and report. Listener 29470 must be GONE in the after-census (no LISTENING row); daemon 39320 stays. If the fetch reads 28: faces share at the wire, STOP, no A4, report and I rule.

Liam: A4 on todlando's A3 stamp only. Hertz: nothing on either box.