Per-worktree index, seeded from a sibling worktree

A git worktree holds a different branch, so a shared codegraph index would answer queries with another branch's symbols. Upstream codegraph calls this the "borrowed index" problem and its only built-in fix is a full reindex per worktree (codegraph init -i), which the user rejected as too slow. We decided: every worktree owns its own index, and a new worktree's index is seeded by copying a sibling worktree's codegraph.db (SQLite online backup), followed by one reconcile pass that re-extracts only the files that actually differ on this branch.

Why seeding works

Codegraph stores file paths relative to the project root and keys files by content hash. Git checkout rewrites mtime on every file it writes, so after a seed the reconcile's size/mtime pre-filter lets every file through to the hash check, and only genuinely changed files are re-extracted. The seeded index therefore converges to the exact state a from-scratch build would reach, at the cost of one file copy plus a read/hash pass.

Considered Options

• Per-worktree index with sibling seed (chosen). Per-branch truth, no full rebuild, and it matches upstream's own model: a worktree-local index is the blessed state, so upstream's borrowed-index warning never fires.
• One shared index for the repo (stored outside the trees, e.g. next to the common .git). One build total, but the graph reflects only one branch. Queries from any other worktree return wrong-branch symbols, which the model would cite confidently. Rejected on correctness.
• Symlink .codegraph into new worktrees. Same wrong-branch problem as sharing, plus it breaks the daemon lockfile and codegraph's inode-replacement detection. Rejected.

Consequences

• Sibling discovery is git-native (git worktree list plus the shared common git dir). Worktrees created anywhere - herdr under ~/.herdr/worktrees/, PR checkouts under /tmp/, hand-made - need no special casing.
• Seed source: prefer the main worktree if it has an index, otherwise any indexed sibling; a partial (incomplete) sibling index is accepted because the post-seed reconcile is a full file walk and converges either way.
• The create-or-seed decision happens under codegraph's per-root FileLock, so two concurrent sessions in the same new worktree build once.
• Diverged branches pay "re-extract only the diff". If a repo grows branches that differ by thousands of files, the seed source can be refined to the closest-HEAD sibling; not done until measured.