fix(sync): pump deep-layer cascade deletes through the sync intake

[winlp4ever]

What

S7b of the sync spine work. Migrates the one real oplog bypass onto the submitLocalBatch seam from #259, so a synced board ships deep-layer cascade deletes correctly.

The bug

Deleting a folder whose children live in an unloaded layer runs a two-part cascade (subtree.ts): the loaded layer is removed through the store, and the deeper descendants are swept from the whole-board oplog. That deep sweep recorded its removals straight to the oplog via getBoardPersistenceRef().record(...), bypassing the sync coordinator entirely.

On a synced board the consequence is: the batch lands in the oplog but never triggers a pump, so it ships only opportunistically on the next unrelated edit, and is never serverSeq-stamped promptly. Deep-layer deletes could sit unsent (until a reload / next edit), and a peer wouldn't see the folder's descendants removed.

The fix

After recording, enter the sync-correct intake:

const batch = makeBatch(store, origin, ops)
persistence.record(batch)
await persistence.flush()
getBoardSyncRef()?.submitLocalBatch(batch, { scene: false })

• scene: false because these ops target unloaded layers — they must NOT join the in-scene rebase set (replaying them would inject off-layer nodes into the current scene; the refactor(sync): extract submitLocalBatch as the single local intake #259 review established this).
• On a local board getBoardSyncRef() is null, so nothing changes — the oplog record alone is sync-correct there (no outbox to desync).

No new rebase machinery is needed: because off-scene batches never enter pending, they never reach applyBatch, so the cross-layer inverse-safety concern from the original plan doesn't arise.

Test plan

• New test: on a synced board (sync ref installed), a folder delete with a deeper child routes the sweep through submitLocalBatch exactly once, with scene: false, carrying the deeper node's removal.
• All existing subtree tests pass unchanged — the local-board path (no sync ref) is a no-op, so the deep-sweep behavior is untouched there.
• npm run check-all passes.

Follow-up

• S7c — ADR-SYNC-001 update documenting "one intake, two producers, one seq authority" and the scene-vs-headless rebase split.