refactor(notebook-doc): unified DocChangeset from single patch walk

[rgbkrk]

Problem

Three functions independently call doc.diff(before, after) and walk the same patches:

Function	Crate	Calls doc.diff()?	Produces
diff_cells()	notebook-doc	Yes	CellChangeset
diff_metadata_touched()	notebook-doc	Yes	bool
compute_text_attributions()	runtimed-wasm (local fn)	Yes	Vec<TextAttribution>

The WASM receive_frame() path calls diff_cells + compute_text_attributions = 2 doc.diff() calls with identical patches. The daemon calls diff_metadata_touched separately.

Additionally, the WASM path invalidates metadata_fingerprint_cache on every doc change (including cell source edits and output streaming), forcing re-serialization ~30/sec during execution. See the TODO at runtimed-wasm/src/lib.rs:1440.

Proposal

One function in notebook-doc that diffs once, walks patches once, classifies everything:

pub struct DocChangeset {
    pub cells: CellChangeset,
    pub metadata_changed: bool,
    pub text_patches: Vec<TextPatch>,  // raw splice/delete per cell source
}

pub fn diff_doc(
    doc: &mut AutoCommit,
    before: &[ChangeHash],
    after: &[ChangeHash],
) -> DocChangeset

text_patches

Captures the raw splice/delete operations on cell source text — the same data compute_text_attributions currently extracts, but without actor labels (those come from extract_change_actors via get_changes, a separate query). The WASM consumer combines text patches + actors into TextAttribution.

Consumer changes

WASM (runtimed-wasm):

let changeset = diff_doc(doc, &before, &after);
let actors = extract_change_actors(doc, &before);
let attributions = build_attributions(&changeset.text_patches, &actors);

if changeset.metadata_changed {
    self.metadata_fingerprint_cache = None;  // fixes the TODO
}

Daemon (runtimed):

let changeset = diff_doc(doc, &before, &after);
if changeset.metadata_changed {
    check_and_broadcast_sync_state(room).await;
}

What moves where

• Patch-walking guts of compute_text_attributions move into diff_doc() in notebook-doc
• compute_text_attributions stays in runtimed-wasm as a thin combiner (text patches + actors → TextAttribution)
• diff_cells() becomes a convenience wrapper: diff_doc(...).cells
• diff_metadata_touched() becomes: diff_doc(...).metadata_changed
• extract_change_actors() stays as-is (uses get_changes, not diff)

Benefits

1. WASM: one doc.diff() per sync frame instead of two
2. WASM: metadata fingerprint cache only invalidated when metadata actually changed
3. Daemon: one doc.diff() instead of separate metadata check
4. Extension point: adding new patch categories is just a new field on DocChangeset

Design question

Whether text_patches should surface raw Automerge patch data or a more structured type (cell_id + index + text + deleted). The structured approach is cleaner for consumers but couples notebook-doc to the text attribution concept. Raw patches keep notebook-doc focused on classification only.

Context

• diff_cells and diff_metadata_touched added in refactor(runtimed): patch-based metadata detection and typed EnvKind #1663
• compute_text_attributions is in runtimed-wasm/src/lib.rs:1647
• The metadata fingerprint TODO is at runtimed-wasm/src/lib.rs:1440
• extract_change_actors is in notebook-doc/src/diff.rs:328 (uses get_changes, not diff)