Batched source sync reports complete before earlier items are recallable #7474

Summary

#7253 merged, and with it one behaviour change that was raised in review but never decided: a batched source sync no longer waits for ranked recall on every item, only on the last of each batch. So store_all can return — and run_and_record can mark the source Idle — while earlier items of a batch are listed but not yet recallable.

Problem / Context

Verified in the engine, not inferred:

• cortex/engine/store.rs::store_items ends with await_written(written, index == final_index) per scope.
• cortex/log/write.rs::await_written(_, settle) calls await_readable(...) — the ranked-recall wait — only when settle is true, and otherwise just await_listed(...).
• So exactly one event per batch, the final one, is awaited for ranked recall.
• The engine's own doc is explicit that monotonicity holds for listing and says nothing of the kind for recall: "the scope's log is ordered, so it being listed implies the earlier ones are), and ranked recall for the batch's final event only."
• The path Store memory source items in batches, as backfill already does #7253 replaced went store_on → store_on_with(.., WriteOptions::visible()) → a single-item store_items, where that item is the final event. Every item used to be recall-settled.

There is no stronger batch call to switch to. store_many_on already uses WaitFor::Visible; the settle-for-final-event-only is internal to store_items and not selectable through WriteOptions, and the only other value (WaitFor::Accepted) is weaker still. That is why this is a decision rather than a patch.

What is not at stake: which items get stored, which fail, the is_account_wide() early exit, or the touched node set that keys the belief-build jobs. Listing visibility also still holds, so list-shaped reads (memory_list and friends) see every item when sync reports complete. Only ranked recall can trail, bounded by the index refresh, and only for non-final items of a batch.

Scope (optional)

The question is whether "source sync reported complete" must imply "recallable" on this surface. Three answers, and I do not think an outside contributor should pick:

1. Accept it. Document the trade-off on store_all. import and backfill already live with exactly this, so it would make the three consistent.
2. Revert this path to per-item. Restores the old guarantee exactly and gives back the batching — N network round-trips instead of ceil(N / 100), since the engine is always an HTTP Cortex engine.
3. Ask upstream for the option. A per-item — or last-per-scope — settle flag on tinymemory's store_many_with would keep both properties at one request per chunk. This is the only answer that does, and it is not in this repository.

Acceptance criteria

• Decided — one of the three above is chosen, and the reasoning is written down where the next reader of store_all will find it.
• If (1) — the trade-off is on store_all's doc comment, not only in a PR description, since that is where it merged without being stated.
• If (2) or (3) — the change lands with a test that a sync reporting complete implies every item is recallable.
• Diff coverage ≥ 80% — if the answer involves code.

Related

Raised by review on #7253 (CodeRabbit, Major) and verified against the engine before filing; the PR merged before the question was settled, so it is filed here rather than left in a merged thread. Origin issue: #7252. Related, same domain, already filed: #7366. Follow-up to the same merge: the drain(..n) front-copy in the merged code, fixed separately.