engine: four copies of the spawned-loop poll driver across the integration suites

[FSM1]

The waker-fixpoint driver that pumps the engine's spawned loops is copied per test crate, and the copies have drifted into three different contracts under two names:

• crates/engine/tests/write_plane.rs — WokenFlag + poll_each, polls to fixpoint, returns Vec<Poll<()>>
• crates/engine/tests/vault_settings.rs — Woken + poll_until_parked, polls to fixpoint, returns () (added by PR feat(engine): give the vault settings write plane a facade caller and a renewal slot #1298, renamed to avoid the collision below)
• crates/fuse/tests/fuse_op_core.rs — a third copy
• crates/engine/src/facade.rs's own unit tests — a poll_each that polls once with a no-op waker, which is a different drive model under the same name

engine_on and boot are duplicated the same way between write_plane.rs and vault_settings.rs.

Integration tests are separate crates, so this cannot be deduplicated locally — it will be copied a fifth time, and a fix to the polling contract lands in one copy only.

What to do

Export the driver from crates/engine/src/testkit, beside block_on in testkit/executor.rs — every one of these files already imports from there — and delete the copies. Give the single-poll and poll-to-fixpoint variants distinct names, since the current collision means a reader moving between two files silently gets the wrong one.

Part of #655

---

Triage note 2026-08-19 — overlaps #1121, verify before starting

#1121 already scopes poll_each/WokenFlag, engine_on and boot as part of the published-account fixture it promotes out of write_plane.rs and fuse_op_core.rs, and #1297's "Relationship to #1121" section names poll_each in #1121's list explicitly. If #1121 lands as written, most of this issue is already done.

What is left over and belongs here either way: the Woken/poll_until_parked copy in vault_settings.rs, and the name collision between facade.rs's single-poll poll_each and the integration suites' poll-to-fixpoint poll_each. Preferred outcome is folding that residue into #1121 and retiring this issue rather than running both. Marked blocked by #1121 so the two are not picked up in one wave.

[FSM1]

Retired 2026-08-19 — superseded by #1121, residue folded in first

Retired as not planned, not as done. Nothing here was implemented; the work moved.

What this duplicated. Two of the four copies listed above are already in #1121's promote-and-delete list, named there explicitly: write_plane.rs:662 poll_each + WokenFlag:648, and the fuse_op_core.rs copy — which is spelled pump (:1532) + Woken (:1549), the same poll-to-fixpoint driver under a different name. #1121 also owns engine_on and boot for those two files, and #1297's "Relationship to #1121" section confirms the split, keeping the seam fault stubs separate from this fixture surface. Running both issues would have meant two promotions racing over the same files.

What was genuinely residual, now recorded in #1121's body under "Poll-driver residue folded in from #1303":

• A fourth copy in src/, not tests/ — crates/engine/src/facade.rs:5702 poll_each, a single poll under Waker::noop(). That site appeared in no other issue's list.
• The naming collision, which is worse than reported here. Three distinct drive models sit under four names, and two of the names carry two meanings each, in opposite directions: poll_each is poll-to-fixpoint in write_plane.rs:662 but single-poll in facade.rs:5702, while poll_once polls every task once in write_plane.rs:676 and polls a single future once in fuse_op_core.rs:166. Both collisions are now tabulated in test: promote the published-account fixture into the engine testkit #1121, with distinct names for the two drive models added to its done-when.

Two claims in the body above did not verify at main 8df8f6e03:

1. crates/engine/tests/vault_settings.rs carries no poll driver on main — no poll_until_parked, no Woken, no poll_each. The name poll_until_parked exists nowhere in the tree.
2. engine_on and boot are not duplicated between write_plane.rs and vault_settings.rs on main; vault_settings.rs has neither.

Both are true only of PR #1298's branch, which is unmerged. That copy is real and incoming, so it is recorded in #1121 as a rebase-time sweep rather than dropped.

Dependencies. Nothing was blocked by this issue, so retiring it orphans no dependents. It was itself blocked by #1121; that edge carries no meaning past the retirement, and the direction it encoded is now moot, because the work sits inside #1121 rather than waiting on it.

Pick the poll-driver work up in #1121.