todlando — WINDOW START 2026-09-11T10:20Z. Box taken on hertz's window 3 END (10:17:14Z). ONE producer, ONE attempt, no retry, cap 2400 s.
SUBJECT: COMMITTED blob 0845e68d, feat/304-remote-friction-product, worktree .worktrees/304-product. Tree clean but for the ADR-0058 Amendment 1 draft. Pool .worktrees/304-product/target, label todlando-304-product. Build jobs 2.
PRODUCER: cargo nextest run -p spt-daemon --test attach -E test(=admission_answers_the_clients_own_identity_with_applicable_policy). Already compiled and enumerated under the granted compile-only leg (16 population / 1 selected, named_but_absent empty), so this leg is execution.

ADMISSION 10:20:01Z: free 179,231,158,272 bytes (166.92 GiB), far above my 96 GiB floor. Stop at <= 32 GiB free or >= 64 GiB growth.
CENSUS 10:20Z, and the path-keyed arm now uses hertz's reading rather than my own: psutil p.exe() instead of CIM ExecutablePath, and an EMPTY path counted as blind rather than skipped silently — his two corrections, both adopted, neither on faith.
  total processes 558 · ci=0 agent=0 unknown=0 analyzer=4 (26640/46072 rust-analyzer.exe, 39260/43280 rustup.exe shims) · pool-deps executions 0
  BOTH controls TRUE: one process wearing both keyed properties (copied to the name cargo.exe AND launched from the pool's debug\deps), alive across the poll, pid 30800, killed after, residue verified ABSENT.
  BLIND 3 of 558: pid 0 System Idle Process (AccessDenied), pid 4 System (EmptyPath), pid 236 unnamed (EmptyPath).
CROSS-CHECK WORTH RECORDING: hertz measured 3 of 556 at 09:58Z with his instrument; I measure 3 of 558 at 10:20Z with mine, same three by identity. Two independent readers agreeing on the blind set is what makes it a property of the box rather than of either tool — which is exactly what my CIM number could not tell either of us.
ANALYZER CAVEAT (IR-100), unchanged: those four resolve the WORKTREE manifest, build into MY pool and take MY build lock. Growth is not attributed to me.

PREDICTION: already filed and accepted (doyle, 10:18Z) — arm 1 GREEN; arms 2 and 3 reach their POLICY assertions for the first time with NO COLOUR PREDICTED; structural: five ARM_OPEN/ARM_DONE pairs, five distinct serve_streams, five distinct gens, no timeout, no session-detach before an arm's own echo. FALSIFIER I hold against myself: if arm 2 still reds on the read with distinct input ops, my dedup reading is WRONG and classification reopens. The fourth fix does not get to be the last one by assumption.
END will name run/passed, the arms that actually executed, and every structural check above, whichever way it lands.