Use macrotask-yielding branch of automerge-repo

[expede]

Stacked PR over #38

Benches + LLM Commentary

Offline ingest, single sample, dev box (512 B/file, fanout 20):

files	syncMs 2.6.0-subduction.29 -> feed-macrotasks	maxDriftMs 2.6.0-subduction.29 -> feed-macrotasks
500	1729 -> 2026	551 -> 655
1000	4462 -> 3301	1251 -> 945
2000	13303 -> 7861	2813 -> 2295

feed-macrotasks is faster and blocks less at scale (2000 files: ~41% less wall,
~18% shorter longest block); the win grows with file count (the quadratic-flush
fix). Still multi-second at 2000 — the residual block is partly pushwork's own
pushFiles/folder-encode, not automerge-repo's feed. Gated guard still passes.

---

[Back to me]
This is still prior to the more radical changes to pushwork's underlying strategy from the reliability branch pre-v2.0