test(ci): calibrate the container-build timeout to its CI-measured 95s (INV-CI-4)

[neosun100]

The arm64 build+serve job added in #67 landed with a 20-minute timeout and no entry in _MEASURED_SECONDS, so INV-CI-4's 2×–40× band did not constrain it. Closing that immediately, while the measurement is fresh.

Measured in CI, on a native ubuntu-24.04-arm runner:

runner arch: aarch64        (native — no QEMU)
Docker version 28.0.4
7 passed in 44.64s          (zero skips — the gate really built)
job wall-clock: 95s

Locally, a cold build (Apple Silicon, docker builder prune -af first) is 112s.

20 min against 95s is 12.6× — looser than the ~3× used for the other jobs, deliberately, and the reason is now recorded beside the number: this job pulls a base image and installs the whole specialist stack, so its runtime is dominated by network I/O that a slow registry day can multiply. And unlike the other jobs, a timeout here means the image was never proven to build at all — the failure mode is "we learned nothing", not "we lost a few minutes".

Mutation-tested both directions:

mutation	result
timeout → 1 min (0.6× of 95s — variance becomes a flake)	CAUGHT
timeout → 120 min (decorative, near GitHub's 360 default)	CAUGHT

Full suite: 4085 passed / 15 skipped. ruff clean.