docs(rollout): `RolloutStore::add` doc comment still promises synchronous visibility

[beinan]

Summary

After #181 (feat(rollout): concurrent writes with periodic flush, 8b2c49c), RolloutStore::add no longer seals and drains on the append path — it only does a durable put. The doc comment on add was not fully updated and still describes the old contract.

Evidence

crates/lance-context-core/src/rollout_store.rs:520-542 (doc comment):

Each append is sealed and its flushed generation committed to the shard manifest before returning, so the rows are immediately visible to reads on any instance

The per-append work is put (WAL-durable) → force_seal_active (freeze this append's memtable) → wait_for_flush_drain (await the generation's manifest commit)

The actual implementation at rollout_store.rs:543-576 performs only put. force_seal_active + wait_for_flush_drain now live in flush() (rollout_store.rs:621-640), driven periodically by the server's flush sweeper (crates/lance-context-server/src/state.rs:474-518), default interval 30s (crates/lance-context-server/src/config.rs:58).

The inline comment at rollout_store.rs:549-555 already states the correct new semantics ("Read-after-write is therefore asynchronous, up to the flush interval; durability is not"), so the doc comment directly contradicts the code a few lines below it.

Why this matters

This is a public API doc comment that ships in rustdoc. It tells every downstream caller that a write is globally visible on return, when in the default configuration it may take up to 30 seconds. Durability is genuinely unchanged — only visibility regressed — but the contract as written is wrong.

Proposed fix

Rewrite the doc comment to state:

• add is durable on return (put waits for the WAL entry to reach object storage)
• add is not read-your-write visible; visibility is bounded by the flush interval
• point at flush() for callers that need synchronous visibility

Remove the stale three-step put → force_seal_active → wait_for_flush_drain description.

Notes

Not a correctness bug. Rollout rows are immutable and the LSM read path dedups by id (rollout_store.rs:1690-1732), so the decoupling does not introduce duplicates.