Direct-SSH snapshot tabs are silently dropped when the worktree catalog resolves their paths late #12902

Summary

When a direct-SSH remote workspace snapshot is applied, any snapshot worktree path the local worktree catalog cannot resolve yet is silently dropped and never retried. The tabs under that path — and any sleeping agent sessions bound to them — stay missing until the next full reconnect. To the user this looks like their remote agent session vanishing: activating the worktree shows a brand-new empty terminal instead of the restored session.

Environment

• Orca 1.4.174 (also reproduced on 1.4.171), macOS (darwin arm64)
• Direct SSH workspace to a Linux host, claude agent running in a remote worktree tab

What happens

applyDirectSshRemoteWorkspaceSnapshot (src/renderer/src/hooks/remote-workspace-snapshot-apply.ts) imports the pulled snapshot through importRemoteWorkspaceSession with resolveWorktreeId: uniqueWorktreeIdByPath(worktreeIds). In importRemoteWorkspaceSession (src/shared/remote-workspace-session-projection.ts):

for (const [worktreePath, tabs] of Object.entries(remote.tabsByWorktreePath ?? {})) {
  const worktreeId = resolvePath(worktreePath)
  if (!worktreeId) {
    continue          // <-- tabs for this path are dropped, no retry ever happens
  }
  ...
}

The resolver is built from the worktree catalog at apply time. For remote hosts the catalog populates asynchronously — in practice a remote worktree often only resolves after the user activates it in the sidebar. The snapshot apply, however, runs once on connect/reconnect. So the ordering is:

1. SSH connects, snapshot is pulled and applied. The catalog does not know the remote worktree path yet → its tabs are dropped from the import.
2. The sleeping-agent wake sweep finds its resume record but logs tab <id> ... not present in any worktree (likely pre-hydration); skipping forever — the tab never appears.
3. The user clicks the worktree. The catalog now resolves the path, but nothing re-runs the import. The worktree hydrates with no tabs, Orca creates a fresh empty terminal tab.
4. A subsequent workspace export pushes that tabless state back to the host, so the snapshot itself loses the tabs — the session loss becomes durable.

Reproduction

1. On a direct-SSH workspace, run an agent (e.g. claude) in a terminal tab of a remote worktree; let the workspace snapshot sync.
2. Quit Orca. Kill the remote relay daemon on the host (simulates relay restart/update — same thing happens after an app update deploys a new relay).
3. Relaunch Orca and, before activating anything, let SSH reconnect and the snapshot apply run (catalog is still cold).
4. Activate the worktree: it comes up with a new empty terminal; the snapshot tabs and the agent resume record never fire.

Expected

Snapshot tabs whose worktree path resolves later should still be hydrated — the import should be retried (bounded) as the catalog fills in, so restored sessions attach to their original tabs instead of being lost.

Fix

PR incoming from this account: it appends a bounded deferred-hydration watcher to applyDirectSshRemoteWorkspaceSnapshot that re-imports late-resolving paths scoped to just those worktrees, additively (never removing live tabs) and without moving user focus. Running patched locally; recovers the exact scenario above.

Related: #12722, #12704, #12751 (adjacent restore/merge issues in the same flow).