D20-4 reproduction, 2026-11-08, this Mac (the test machine's macOS), copies of the installed app's llama-server. Runs: <binary> --help, stdout piped.
Scripts: t3.py, t4.py (limit 5 s = local_models.rs:288).

Installed app, first run since install (/usr/bin/time): 11.99 s; next two runs 0.57 s
Fresh copy A, no limit: 12.50 s; again 0.58 s
t3 new parent 1, limit 5 s: KILLED at 5.01 s, no output before the kill; retries 1-3: KILLED at 5.01 s each
t3 new parent 2 (TMPDIR), limit 5 s: 0.84 s; retries 0.61 s
t4 copy 1, limit 5 s: KILLED at 5.01 s; retry limit 5 s: KILLED at 5.00 s; no limit: exit 0 after 12.04 s; after that, limit 5 s: 0.59 s, 0.59 s

Not every fresh copy is slow: copies B-E in t.py and t2.py took 0.88-0.97 s. Which macOS component holds the launch is not established.
Reported by the user on another device: "Local Model · Failed / Runtime flag check cancelled or timed out". The workaround (running --help once in Terminal) was confirmed by the user on that device on 2026-11-08: the error went away and Qwen3.8-27B started.
