[eersnington]

Closes #20.

Overlapping ctx.join branches could save history that loadStorage cannot read, so replay failed with Name index N not found in registry. On actors, overlapping saves also failed with State cannot be mutated by another action while an experimental state-aware database transaction is active. Each failure counted as a step failure, so finished steps ran again and large joins ran out of retries.

• flush() runs one save at a time per workflow, so an older snapshot cannot land after a newer one.
• Saves that have not started share the next write. Saves with deletions keep their own write, so pruning stays in order.
• A save marks only the names, entries and metadata it wrote as saved. Changes made during the write stay dirty.
• A failed write or delete marks its data dirty again, so a later save retries it.
• executeStep marks step metadata dirty when a step finishes or fails, so a sibling's earlier save cannot leave it as running.
• Second commit, to squash: formats examples/docs/onboarding/index.ts so format:check passes, and removes the dangling docs/AGENTS.md symlink (its target was removed in baba281).

20-branch join on actors, rivetkit 3.0.4	main	This PR
Finishes, no added latency	not within 60 s	121 ms median
Finishes, 50 ms round trip	not within 60 s	414 ms median
Steps run twice with no fault	66 of 190	0

Same fix in engine-ee: rivet-dev/engine-ee#1776.

---

This is part 1 of 5 in a stack:

•  5  fix(workflows): stop long step timeouts from firing immediately #28
•  4  fix(workflows): wake race branches at their deadlines while other branches run #27
•  3  fix(workflows): reject removed try scopes #26
•  2  fix(workflows): mark settled history as visited on replay #25
•  1  fix(workflows): save overlapping join history in order #24 👈