Merging user storage drops non-managed files from the losing conversation copy

• Status: Todo
• Kind: Bug
• Authors: jp
• Date: 2026-09-27
• Label: domain=storage
• Label: package=jp_storage
• Label: type=bug

When the same conversation exists on both sides of a user-storage merge, the copy that loses the mtime comparison is deleted whole, so a query draft (QUERY_MESSAGE.md) or any other file kept only in that copy is lost without a diagnostic.

Why

adopt_conversation_dir (crates/jp_storage/src/lib.rs) keeps whichever copy has the newer dir_mtime:

• When the source is newer, it runs fs::remove_dir_all on the destination and moves or copies the source into its place. The destination's non-managed files go with it.
• When the destination is newer, it leaves the source alone. But for a sibling merge (merge_sibling_user_workspace_dirs), the whole sibling directory is removed afterwards, taking the source's non-managed files with it.

Both paths run from migrate_user_storage: the sibling merge on every run where legacy per-worktree directories exist, and the workspace import on a workspace's first run.

Fix

Pick the winner's managed files, not the winner's directory. Bring the destination to the winning name, then carry across the files the loser holds that the winner does not, the way sync_projection copies only MANAGED_FILES and leaves everything else alone. When both copies hold the same non-managed file with different content, keep the newer one and log a warning rather than delete silently.

Verifying

A sibling merge where the losing copy holds QUERY_MESSAGE.md and the winning copy does not: after the merge, the draft is in the surviving directory with its original content. Cover both directions (source newer, destination newer).