What should we build first?

[el-bri-ga]

The plan starts with M0: one node, one workload (#1 to #5), then leases and failover (M1). Before writing code, it's worth checking that order. Some alternatives:

1. Follow the milestones. Workspace and CI, the WIT contract, the deterministic runtime, the resume fold, then a single-node agent. It's the safest order, but nothing visible happens until M1.
2. Failover demo first. Build the thinnest possible path to "kill a node, the workload continues" (parts of M0 and M1 together) and harden it afterwards. It shows the idea sooner, at the cost of some rework.
3. Prove the risky assumption first. Confirm that a publish_if batch is all-or-nothing (Confirm publish_if batch atomicity #10) and measure many single-shard streams (Many workloads: streams per workload at scale #20) before anything else. Exactly-once output depends on both.
4. Replay first. Replay needs no leases or cluster: a runtime, a snapshot and range reads. It's the most distinctive feature and the easiest to show on a laptop.

My lean is 3 then 1: a day spent on #10 could change the design, and everything else builds on M0.

Which would you pick, and why? If you'd like to take one of these on yourself, say so here.