fix(test): settle fake checkpoint writer so main-slice test drains cleanly

[onlyfeng]

Summary

checkpoint-main-slice test: the recording actor's spawn outcome never resolved,
leaving tryStartCheckpointWriter's settlement-watcher fiber suspended until scope
close force-interrupted it — a flaky source under it.live. Settle the outcome so
the watcher completes (writers.delete) and assert the writer drains to no-writer.

No production change — actor/spawn.ts is untouched; the test's spawn-ref finalizer
stays = undefined, consistent with the sibling checkpoint tests.

Verification

• bun typecheck — exit 0
• bun test test/session/checkpoint-main-slice.test.ts — 1 pass / 13 assertions

[onlyfeng]

Note for reviewers — the failing unit (shard 4/4) checks are pre-existing flakiness on the base, not introduced by this PR.

• The two failing tests (CheckpointContext producer ... after settle and parentSessionID end-to-end ... splitover) already fail on the base commit 0abfd8a on main (push CI at 2026-06-27 02:30), before any change here.
• This PR's only change is test/session/checkpoint-main-slice.test.ts; src/actor/spawn.ts is untouched. Locally these all pass: checkpoint-main-slice, checkpoint-splitover-integration (standalone and combined with checkpoint-main-slice, both orders).
• Root cause looks like shard-level test isolation: under shard load, global spawnRef/isSystemSpawned state gets perturbed and tryStartCheckpointWriter takes the "skipped" branch instead of "started".