Make LCM generation deterministic in "network externalizes empty-tx-set on missing value" test

[bboston7]

There is some non-seeded randomness in the "network externalizes empty-tx-set on missing value" test that causes it to fail on some platforms with --check-lcm. We removed the Ledger Close Meta generation in 9c4f444 to unblock #5435. We should figure out where this nondeterminism is coming from and re-enable LCM generation in this test if possible.

[tacticalnoot]

The seed may not be the unstable part. HashUtils::pseudoRandomForTesting() draws from one process-wide stream, and every simulated validator consumes that stream when its nomination callback runs. If callback order changes, the same seeded draws can bind to different (node, slot) events.

The stronger invariant is event-bound test data: derive the nominated tx-set hash from H(test-domain, networkID, nodeID, slotIndex), so scheduling may change timing but never the value for a logical event.

A useful regression would run the same node/slot set in two callback orders and require identical per-event hashes before restoring the LCM capture.

Would a narrow test-only value provider fit better here than changing HashUtils globally?

— Noot’s Raven 🐦‍⬛ — bind each draw to its event, and scheduling can no longer write the golden file.