M1-03D - socket load check and bench re-runs (2026-10-11). Machine: Apple M3, 8 cores (4P+4E), heavily loaded by
other work: load average 86-106 during the load check, 58-103 during the benches (load-average.txt).

pnpm zone:loadcheck (new; services/zone/src/bin/loadcheck.rs) - 200 WebSocket clients, 10 s, in-process server on an
ephemeral loopback port (loadcheck.json, server-log.txt):
  PASS  200/200 joined, 200 clean 1000 closes answered, server saw 200 leaves, 0 client errors, 0 threads left
  PASS  inputs 36,967 applied, 0 stale, 0 dropped; frames 44,659 sent, 0 skipped; 0 missed ticks seen by clients;
        scheduler 0 dropped ticks, 4 catch-up ticks
  PASS  downstream per client mean 12.8 KB/s, p95 13.8, max 13.9 (hard cap 32 KB/s); snapshot frames mean 604 B,
        p95 777 B, max 1,452 B (packet budget)
  FAIL  loop p99 84.7 ms (target 15 ms); p50 13.6, mean 18.8, max 127.9 ms; 88 ticks over target, 11 over 50 ms
  info  tick-thread CPU 6.43 ms per tick = 34 % of the loop's wall time: the tick spent about two thirds of each
        iteration waiting for a core. The p99 miss is machine load; per-tick CPU p99 is not measured (mean only).
  info  upstream per client 521 B/s (doc estimate 0.3 KB/s counts inputs only; acks per snapshot and pings add the
        rest); RTT p50 41.6 ms, p99 62.5 ms (pings are answered by the tick, so ~1 tick of server queueing).
  Earlier exploratory runs at load 60-80: 50 clients p50 1.3 ms, 100 clients p50 12.6 ms; 20 clients at load 64:
  tick CPU 0.12 ms/tick but loop mean 7.2 ms (CPU share 1.7 %) - the same load signature.

pnpm zone:snapbench re-run (snapshot-bench.json, snapbench-excerpt.txt): bandwidth identical to M1-03B and within
  budget (crowded town max 10.4 KB/s); one-thread tick+snapshots p99 47.6-241.2 ms, still not an idle number.
pnpm zone:bench re-run (tick-bench.json, bench-excerpt.txt): PASS, p99 7.6 / 9.5 / 12.0 ms on one thread, 0 allocations.

Reading for M1-04: everyone-in-one-spot at 200 clients costs ~6.4 ms of tick CPU (mostly snapshot building, going by the tick bench). A 1,000-player
crowd on one thread would not fit 15 ms p99; NETCODE_1000.md already plans parallel per-client work over 4 cores.
An idle-machine re-run of `pnpm zone:loadcheck` is still owed (fold into the M1-04 soak run).