boot: reconcile fast-forward does not re-open the store snapshot, so in-memory state is built from the pre-reconcile tree

[themightychris]

What

buildApp registers the plugins in this order: store (opens the gitsheets public store; each Sheet caches a dataTree snapshot at open time) → reconcile (fetch + fast-forward/rebase against origin/<CFP_DATA_BRANCH>) → services (builds InMemoryState + FTS from fastify.store.public).

When the local bare clone is behind origin at boot, reconcile fast-forwards the branch, but nothing calls Store.swapPublic afterwards. services then builds the in-memory state from the Sheet snapshots captured before the fast-forward. Records that arrived in the fast-forward are invisible until the next hot-reload webhook or a restart.

The hot-reload path (reloadInMemoryStateAndFts) already handles this correctly by re-opening the public store after reconcile. The boot path skips that step.

Why it hasn't bitten

Production pods bare-clone the data repo on every boot (emptyDir volume), so the clone is in sync with origin by the time reconcile runs and the outcome is in-sync. The gap only shows when a clone is reused across boots: local dev, and tests that seed the remote after creating the rig (see the re-import test in apps/api/tests/internal-reload.test.ts, which works around it with an explicit git fetch origin main:main before boot).

Fix sketch

In the reconcile plugin (or a small step between it and services), when the outcome is anything other than in-sync/fetch-failed, re-open the public store and fastify.store.swapPublic(freshPublic) before services reads it. Or reorder so the store opens after reconcile; reconcile only needs the repo path and the lock, not the Sheet handles.

Found while working on the hot-reload stale-indices fix (fix/hot-reload-stale-indices); out of scope there.