feat(rollout): expose a synchronous-visibility write option on the HTTP API

[beinan]

Summary

Since #181, a rollout append is durable on return but not visible until the next flush tick (default 30s). RolloutStore::flush() exists (crates/lance-context-core/src/rollout_store.rs:621-640) but is not reachable from the HTTP API, so callers that need read-your-write have no way to get it short of setting the global flush interval very low.

Motivating cases

• an RL training loop that appends a trajectory and immediately reads it back to verify
• write-then-assert in integration tests
• any client doing POST /rollouts/{name}/records followed by GET

All of these silently observe a missing row for up to rollout_flush_interval_secs.

Proposal

Add an opt-in per-request flag on the rollout append route (crates/lance-context-server/src/routes/rollouts.rs), e.g. ?visible=true or an X-Rollout-Visible header, that performs add() followed by flush() before responding.

• default stays asynchronous, preserving the concurrency win from feat(rollout): concurrent writes with periodic flush #181
• the flag is explicitly documented as trading latency for immediate visibility
• consider surfacing it in the Python binding as add(..., wait_visible=True)

Alternative considered

Lowering the global flush interval affects every store in the process and re-serializes writes that do not need the guarantee. A per-request opt-in keeps the choice with the caller.

Related

• feat(rollout): concurrent writes with periodic flush #181, docs(rollout): RolloutStore::add doc comment still promises synchronous visibility #183