Re-sync the Notebook panel after notebook.md writes that bypass the watcher

[dannon]

Closes #253

The Notebook panel only refreshed when the chokidar watcher caught a change to notebook.md, but the watcher tracks a single file path. An atomic temp-and-rename save (sed -i, an editor) swaps the inode out from under it, and a plain truncate can race its awaitWriteFinish window -- so when the agent wrote the notebook through bash instead of /notebook, the panel went stale. The tool_execution_end hook also only re-emitted for galaxy_ tools.

What this does

tool_execution_end now calls a new reemitNotebookIfChanged() (in state.ts) after every tool. It re-reads notebook.md and emits through the existing notifyNotebookChange path only when the content actually changed:

• a dev:ino:size:mtimeMs:ctimeMs stat fingerprint skips the read when the file is untouched, so it isn't I/O churn on every tool end, and
• a content comparison suppresses no-op emits.

Including ino/ctime in the fingerprint (not just mtime+size) is deliberate: an inode-replacing or same-size rewrite with a preserved/coarse mtime is exactly the watcher-evading write this is meant to catch.

Two small hardening changes fell out of review of the new emit path:

• notifyNotebookChange now isolates per-listener exceptions (logged, not rethrown), so one misbehaving subscriber can't break the tool_execution_end hook or starve other listeners.
• setNotebookPath resets the emit baselines, so switching to a new notebook that happens to share the old one's content/fingerprint still re-syncs the panel instead of being suppressed.

Scope

This is the contained UI re-sync fix only -- the watcher itself is unchanged (parent-dir watching / a watcher rewrite is a separate, larger change). Two deliberately out-of-scope notes:

• Duplicate emit in the race window: if reemit fires before the watcher's 500ms awaitWriteFinish settles, the watcher later emits the same content again. The only onNotebookChange listener (the UI bridge) dedups on content, so there's no visible double-render.
• Auto-commit: a write the watcher truly misses refreshes the panel but isn't auto-committed (the watcher owns commits). That's a durability concern for the watcher work, not this UI fix, and it's not a regression -- previously such a write neither refreshed nor committed.

Tests

New tests/notebook-reemit.test.ts (7 cases): changed -> emit once, unchanged -> no emit, no re-emit on a second call, no-op without an active notebook, an inode-replacing write with pinned size+mtime still emits (fails on a mtime+size-only guard), a throwing listener neither breaks the caller nor starves other listeners, and a path switch to a same-content notebook still re-emits. The cases run with the real watcher stopped so they're deterministic. Full root suite green (1257 passing); root and Orbit typechecks clean.