auto_watch misses successful index_repository targets from another session root #2733

With auto_watch=true, an active MCP session can successfully index another repository or linked worktree, yet that target never joins the background watcher. This also affects explicitly named indexes.

Reproduction:

1. Start an MCP session in an already indexed workspace.
2. Call index_repository(repo_path="/workspace/service-worktree", name="service-review", mode="fast") and wait for success.
3. Call index_status(project="service-review").
4. Edit a source file in that worktree.

The target reports watched: false / not_session_project and remains stale until explicitly reindexed. The session coordinator registers only its path-derived startup project and stores one watch membership per session.

Expected behavior: with the existing auto_watch setting enabled, successfully indexed targets should be owned by the requesting live MCP session as additional watches. Keep their actual project names and root paths, retain other session watches, and release a watch when its final owner disconnects. Async indexing should attach only after success and only if the requester is still connected. Failed indexing and one-shot CLI calls should not acquire watches.

This is the automatic registration slice related to #1429 and #1660. It preserves session-scoped lifetime and the existing MCP tools/parameters; it does not require permanent daemon-owned watches.

Prepared with Codex on behalf of wargloom.