RemoteGitRef.Tree kept a separate snapshot metadata cache keyed by the resolved ref, remote URL, checkout depth, tag inclusion, and discardGitDir option. That path searched git-snapshot metadata before doing a checkout and reopened the indexed snapshot directly by snapshot ID. This was a conceptual carryover from the previous implementation model, separate from dagql's own result retention and pruning semantics.

That metadata index was not a dagql retention edge. When dagql pruning removed a persisted GitRef.tree result, it correctly released the Directory snapshot lease and containerd could delete the snapshot, but the git-snapshot metadata entry could remain visible. A later GitRef.tree evaluation then found the stale metadata entry and failed with the removed snapshot key instead of treating the pruned result as a cache miss.

Remove the core remote-git git-snapshot fast path entirely. GitRef.tree is already a persistable dagql field, so dagql is now the only persistent cache authority for this core API path. If the dagql result is still retained, the resolver is not run. If dagql has pruned it, the resolver performs a fresh checkout and creates a new Directory snapshot.

The remote git mirror lock remains, but it now uses a local lock prefix instead of the deleted cacheref helper constants. The separate legacy git source code still has its own git snapshot metadata path; this change is intentionally scoped to the core remote GitRef.tree resolver.

Verified with:

go test ./core -run '^$'
go test ./core/integration -run TestEngine -count=0
dagger --progress=plain call engine-dev test --pkg ./core/integration --run='TestEngine/TestLocalCachePruneRemoteGitSnapshot' --timeout='30m'
dagger --progress=plain call engine-dev test --pkg ./core/integration --run='TestGit/TestCaching' --timeout='30m'

Signed-off-by: Erik Sipsma <erik@sipsma.dev>

Remote git tree results can be materialized concurrently by different clients before dagql has a completed equivalent tree result to reuse. Those independent checkouts have identical Git content, but worktree files get checkout-time mtimes. Snapshot diff checks metadata before falling back to content comparison, so equivalent checkouts can report the whole worktree as modified.

Normalize every checked-out path to the Unix epoch before committing the checkout snapshot. Use UtimesNanoAt with AT_SYMLINK_NOFOLLOW so symlink entries are handled directly instead of following their targets. This keeps the change local to checkout materialization and leaves dagql/egraph identity unchanged.

This is intentionally only the worktree timestamp normalization step. With keepGitDir=true, .git metadata can still differ between independent checkouts and needs separate normalization.

Signed-off-by: Erik Sipsma <erik@sipsma.dev>

Remote git tree checkouts can still differ after normalizing worktree mtimes when keepGitDir=true. The worktree file contents match, but Git's index records checkout-local stat metadata for tracked paths. Independent checkouts of the same ref can therefore produce different .git/index bytes, which shows up as a .git diff even though the repository content is equivalent.

Run `git read-tree HEAD` after checkout and submodule update when the .git directory is retained. This rewrites the top-level index from the HEAD tree without updating worktree files, because we intentionally do not pass `-u`. The logical index still represents HEAD, but the checkout-local stat cache entries are cleared so independently materialized checkouts converge.

This pairs with checkout mtime normalization: read-tree makes the index content deterministic, then the timestamp normalization pass makes the index file metadata deterministic along with the rest of the checkout. The result keeps the fix local to git checkout materialization and avoids reintroducing the separate snapshot metadata cache that was redundant with dagql.

Signed-off-by: Erik Sipsma <erik@sipsma.dev>