Historical source and lifetime contract

A logical content UUID identifies bytes, not a representation in the current workspace catalog. Selected-tip publication can retain an ancestor's immutable snapshot while omitting that ancestor's payload metadata from the current DB.

repo/source-view resolves one selected source to an immutable DB and a pure tree. It is a short-lived reader, not a retention root. The caller keeps its store and physical ancestry alive while using it. Tree entries contain only logical content IDs, sizes, and modes; no live handles or provenance are stored. repo/read-entry accepts a third source argument for historical entries. Its two-argument form reads the current catalog. repo/read-at resolves the source and path together. CLI diff/log/show retain separate before and after views. Structural diff output avoids payload reads unless rename similarity or whitespace filtering requests them; raw IDs, patches, and statistics need bytes.

Checkout, restore, and mixed/hard reset validate paths and dirty-state rules, then restore missing selected bytes through the existing guarded content writer before their public installation transaction. Logical content IDs remain the same even when preparation chooses a full representation instead of a delta or chunks. Merge uses the same helper. A later preparation/refusal can leave immutable content in the private catalog; HEAD, worktree, and index installation remains one transaction under the caller's serialized workspace ownership. Soft reset does not prepare content. Selected historical publication/advance reads representation closure from the selected snapshot and logical ancestry from the captured catalog. Existing serialized target guards still apply.

Datahike 0.8.1910 retention

Supported retention requires :commit-graph? true, :keep-history? true, retained native branch/merge ancestry, and no history-pruning cutoff. The pinned default d/gc-storage uses the beginning of time as its history cutoff. File-store regressions remove the publishing branch, close all connections, reconnect, collect with that default history cutoff and a zero sweep floor in the isolated fixture, reconnect again, then reconstruct actual full/delta/chunk historical bytes, delta-base content, and exact packed bytes through their index shards. These prove more than snapshot-record existence.

A logical :geschichte.commit/snapshot UUID is an ordinary attribute, not an independent collector root. A restrictive remove-before can stop native-parent traversal and remove that snapshot even while its logical commit row remains. The regression characterizes this unsupported configuration explicitly. Do not run such pruning when promising retained logical history without separately establishing supported physical roots and their lifecycle. This change chooses no new root integration or reclaim policy. Collection during active preparation, shared/operator collectors, external roots, and process-death cleanup need the host's ownership and collector policy; they are not proven by these tests.

Selective logical publication excludes private work/index/config/refs from the current canonical catalog. Its native merge ancestry may retain private physical history for snapshot lifetime. It does not promise selective physical retention.

Semantic decisions left explicit

The existing merge-base policy minimizes combined shortest-parent distance, then logical-ID order. A redundant shortcut edge can make it choose an ancestor of another common ancestor. The minimal DAG regression records that behavior; it is not Git's best-common-ancestor antichain. A future behavior change should select the antichain under the graph budget and choose an explicit multiple-base policy, for example refusal before implementing recursive virtual bases.

Authored logical parents are a set. Bounded authored Git projection sorts parent OIDs for stable cross-peer identity, so it does not preserve semantic first-parent order. Imported exact objects retain their original serialized order. Adding ordered-parent metadata requires a versioned authored projection and must never rewrite already acknowledged OIDs. These decisions are separate from historical source repair.