Review: approved ✅

This is the PR that stops the protocol work from forking session history into a parallel universe, and it takes the conservative route at every choice point.

What I checked

• Same id in both places. A protocol thread materializes into canonical session-v1 JSONL under the same id, so existing CLI and desktop session indexes stay continuous. Users don't see their history split by which client created it.
• Rich lifecycle snapshots stay in threads-v1, separate from the canonical stream. Right layering — canonical stays the interoperable format, protocol-specific richness doesn't leak into it.
• Legacy import is lazy, on first protocol resume, as completed compatibility turns. Lazy means no big-bang migration pass that could damage history; "completed compatibility turns" means imported content can't be mistaken for a resumable live turn.
• Uses the shared cross-process writer lock from Sessions: canonical v1 writer and cross-process ownership #185 rather than inventing a second one. This is the detail that would have quietly broken concurrent CLI + desktop access if it had been missed.
• Preserving titles written by compatible clients avoids the annoying regression where resuming a session through a new path blanks its name.
• The idempotent SessionManager materialization boundary matters because resume can be retried; a non-idempotent version would duplicate turns on a reconnect.

Validation: CI green; full suite green locally at the stack tip.