[vladwing]

Callers lookups miss callers at some commits a --git run indexed, and the call_edges row count differs between identical --git runs.

sync_for_files() deletes every call_edges row of each path in a batch, at every version, before appending the batch's edges. A path has one version per commit that changed it, and parallel inserters write those versions in different batches, within one run or across runs. Each batch therefore deletes the edges of the path's other versions, which callers lookups need at the revisions that hold those versions. Which versions survive depends on batch order, which is why the row count varies.

sync_for_versions() replaces sync_for_files() and deletes only the versions the batch writes, with one IN list over the key concat(caller_file_hash, ':', caller_file). The key is exact because a git blob id holds no :.

A process-global lock now serializes call_edges writes, as content upserts already are. Without it, concurrent deletes on one table conflict and retry, and two syncs of one version could both delete before either appends, leaving its rows twice. The backfill runs only on an empty table and appends without deleting.

tests/call_edge_versions.rs fails on main and passes here. It indexes c0..c1 and c1..c2 as two process_git_range() calls and checks the callers at each commit. The same-version concurrency unit test fails on every run with the lock removed.