nix: record the measured integration VM boot profile

Boot spike: 18.7s total, userspace critical chain is dhcpcd -> network-
online -> docker, i.e. the remaining cost is functional, not trimmable.
microvm.nix evaluated and rejected (boot is ~10-15% of image-load + test).
Replace the assumed numbers in the comment with the measured ones.

Author: kradalby

nix/tests/integration.nix

@@ -75,12 +75,18 @@ in
     networking.dhcpcd.denyInterfaces = [ "veth*" "br-*" "docker*" ];
 
     # Trim the guest toward a tiny profile: this VM only needs to boot, run
-    # docker, and exec one test binary. Boot is ~20s and dominated by dockerd
-    # init (functional, can't trim) plus default NixOS userspace; dropping doc
-    # generation and unused services shaves the avoidable part. (A microvm.nix
-    # guest would go further but still needs a full docker daemon + nested KVM,
-    # so it is a larger follow-up than this profile for a small boot win against
-    # the ~2-3 min image-load + test runtime.)
+    # docker, and exec one test binary. Measured boot is 18.7s (1.7s kernel +
+    # 3.9s initrd + 13.1s userspace); the userspace critical chain is
+    # dhcpcd (6.1s) -> network-online.target -> docker.service (+3.5s), i.e. the
+    # remaining cost is functional (docker waits for the network), not fat to
+    # trim. documentation + timesyncd off shave the cheap, avoidable part.
+    #
+    # microvm.nix was evaluated against this and rejected: it still needs a full
+    # docker daemon + nested KVM, and boot is only ~10-15% of the 2-3 min
+    # image-load + test runtime, so a new flake input can't pay itself back. The
+    # one real lever left, dropping docker's network-online wait (~6s), sits
+    # right next to the dhcpcd interface races handled below and isn't worth
+    # destabilising the matrix for.
     documentation.enable = false;
     services.timesyncd.enable = false;