---
name: bench-wiring-landed
description: "Post-v0.52.0 operator directive complete 2026-08-02 — #108 bench-wrap wiring + scoped CARGO_INCREMENTAL=0 on main @bc58712; twohost red triaged to network window via declared three-arm count read."
metadata: 
  node_type: memory
  type: project
  originSessionId: a6044c46-9fd0-4ab0-831a-cea2fc0b2b7d
  modified: 2026-08-03T00:44:36.953Z
---

Operator directive (2026-08-02, post-v0.52.0) landed: main ff'd 32a8bad → **bc58712** (hertz lane, golden-proved twice at same sha). (1) bench-wrap wired into golden.yml — 8 linux + 7 windows + 3+3 twohost wrapped steps, HEAVY filter hoisted to job-level env (single copy; standing check = per-phase tests-run counts, baseline linux 2584/175, windows 2605/190, reproduced twice). (2) CARGO_INCREMENTAL=0 applied ONLY to the genuine cold build (n1-gate pinned old-broker cache) + local rig recipes (docs/GOLDEN-CI.md); warm legs untouched pending [[warm-delta]] measurement (filed releases#127).

First run's twohost red: B's E-1 send hit a 10s QUIC bound, A cascaded (15-min convergence wait on a reaped peer). Resolved by declared three-arm count read on a same-sha quiet-predicate rerun — arm 1 fired exactly (PUMP_PEER_FAIL a 0→3→0, b 8→22→8; connect-timeouts a 0→3→0): network window on the wifi-only box, not the wiring. The 55s no-listener window (A's pre-relink reap → first daemon of next rung) stays NAMED-NOT-ADOPTED. RTT-probe predicate axis filed as releases#126.

**Why:** the declared-arms-before-dispatch pattern turned an unreadable n=1 rerun into a discriminating experiment; counts-beside-results made the HEAVY hoist provable by repetition.

**How to apply:** unregistered golden failure ⇒ no auto-rerun; rerun only with quiet predicate + a read declared for every arm INCLUDING the "predicate was wrong" arm. Related: [[swarm-discovery-gc-spin]], [[serial-legs-share-runner-workdir]], [[shell-bash-adds-pipefail-sigpipe]].