Text edits previously recorded ['set', path, whole_value] twice per
keystroke (ops + inverse_ops), so history stored a full copy of the
paragraph per keystroke and the op stream carried no meaning beyond
"content changed".

Transaction.splice(path, start, delete_count, text) records
['splice', [node_id, prop], start, delete_count, text] instead.
Mark/annotation range adjustment happens inside op application
(splice_annotated_text in doc_utils.ts, the single source of truth used
by forward and inverse ops alike), and range nodes whose range collapsed
away are garbage-collected ref-count-aware, like set() does.

Pure insertions (every keystroke) take a fast path: insertion
adjustment can never remove or truncate ranges, so the inverse is a
plain splice by construction and no value copies or round-trip checks
run on the hot path. Deletions verify losslessness with a round-trip
check; lossy deletions carry the prior ranges verbatim as a 6th op
element, so undo is always exact.

insert_text and delete_selection (text branch) now emit splices; their
public behavior is unchanged, and plain typing returns early before the
mark/annotation restoration copies. History batching coalesces
consecutive contiguous typing splices into one op using the inverses'
stored lengths (no re-segmentation of the accumulated batch text), so a
one-second typing batch stores a single small op instead of one string
copy per keystroke.

Character offsets are grapheme clusters (Intl.Segmenter), the same unit
as all existing svedit text offsets.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>