Migrate to 1.0

Tantra 1.0 uses a process-wide actor Runtime and fresh storage. It does not read pre-1.0 logs or recover in-flight work across processes.

Before 1.0	1.0
create_session(...)	Runtime.create(...), returning a UUID
synchronous coordinator invocation	Connection.prompt(...) returns TurnResult
merged execution streaming	Connection.send(...), then iterate the Connection or Runtime.events(...) independently
replay(...)	Runtime.events(agent_id, after=seq)
process-loss recovery	answer a live ask on the current writer; after interruption, use a later send(...)
blocking ctx.spawn(...) and ctx.fan_out(...)	model tools spawn, send, and finish
one merged descendant stream	subscribe to each actor journal with its own scalar cursor
cross-process session claims	one root tree per Runtime process; newest local writer connection wins

Every writable Connection mutation now requires a UUID command_id: send, prompt, answer, and cancel. Runtime.create does not. Save the last integer seq separately for each subscribed actor.

Start applications on empty 1.0 storage. A process crash does not restart model calls or tools. A later root send marks an unmatched active turn interrupted and drains only commands that never started.