[raphael]

Problem

streaming.Stream.Snapshot returned ErrStreamDestroyed for any stream that had been destroyed and then created again, when the caller read through a new handle. Writes to that stream kept succeeding, so a caller saw a stream whose sequence numbers advanced while every snapshot came back empty.

Where it showed up: Chorus deletes a channel's stream when the channel's last member expires. A channel used again afterwards lives in stream generation 2. Chorus reads each channel through a new handle, and treats ErrStreamDestroyed as "nothing retained". So once the channel was recreated, read, watch and the dashboard all showed zero messages.

Cause

A recreated stream stores its events under a generation-scoped key, such as <key>:generation:2. A new handle is unbound: it has no generation yet and still carries the generation-1 key until its first operation binds it.

Snapshot's Lua script compared the handle's key with the stream's current key even when the handle was unbound. The publish (AddOnce) and replay scripts already check the generation and key only for a bound handle; the snapshot script was the one exception.

Change

The snapshot script now uses the same condition as replay: it checks the generation and the physical key only when the handle is bound. A handle still bound to a destroyed generation is still rejected.

TestSnapshotReadsRecreatedStreamFromNewHandle covers the case:

1. destroy a stream and create it again (generation 2);
2. read it from a new handle, which returns the new event and binds to generation 2;
3. read it from a handle still bound to generation 1, which returns ErrStreamDestroyed.

Without the fix, step 2 fails with the exact error seen in production.

The full race suite and go vet pass. The change has no wire, key or format change, so mixed versions are safe.