fix(tui): stop one bad event from dropping the rest of a flush batch

[johnnyeric]

What

Fixes the TUI stall where a turn visibly finishes (footer shows the completed duration) but the prompt keeps spinning and the transcript stops updating.

Why

The SSE client batches events and flush() clears the queue before iterating, emitting each event unguarded inside a single Solid batch(). The legacy message.removed / message.part.removed handlers in sync.tsx called search(...) on store.message[...] / store.part[...] without checking they exist — for a session this TUI never hydrated (background-session cleanup, revert, another window), that is undefined and search throws reading items.length. One such event mid-batch permanently dropped every later event in that flush, including the idle status the spinner depends on.

Changes:

• sync.tsx: guard the legacy message.removed / message.part.removed handlers with the same missing-collection checks the .1 handlers already have, and handle session.idle by resetting session_status to idle.
• sdk.tsx: isolate event dispatch per handler and per event so one throwing handler logs (tui event handler failed) instead of dropping the rest of the batch or killing the SSE loop.
• Tests reproduce both symptoms on the existing sync fixture; both fail without the fix.