Incrementally maintained content digest for a window's resident row set (#2658).

The GUI adapter gates each window's content frame-emit on whether the window's rows changed since the last emit. With full-document residence a window's row list is the whole document, so re-hashing the entire list every frame is O(document) and dominates edit-frame build cost. This digest replaces that whole-list hash with a value that updates in O(1) per changed row and stays correct across row insertion and deletion.

Keying

The digest is a fold over per-row cells, where a cell is phash2({row_id, content_hash}). row_id already encodes the row's buffer line (Row.stable_id/4), so a cell captures both a row's identity/position and its rendered content. Two row lists produce the same digest with overwhelming probability when they carry the same multiset of {row_id, content_hash} pairs (see Collision behaviour below for the probabilistic bound).

Algebra

Cells are combined with bxor, which is associative, commutative, and self-inverse. That gives O(1) add/remove/update and makes the digest independent of iteration order (position is already carried inside row_id, so order independence never hides a real change). Editing a row back to its prior content restores the prior cell and therefore the prior digest, so no spurious frame-emit occurs.

Collision behaviour

The digest is phash2-derived, so a single-row content change is missed only when the old and new cells collide (~1 in 2^27 per changed row). For multi-row splices the deltas of several rows can also XOR-cancel even when no single row collides; the overall miss probability stays in the same ~2^-27 order. This matches the collision profile of the whole-list phash2 it replaces; it does not add risk. Row ids are assumed unique within a window (guaranteed by Row.stable_id/4 per position); the digest is a set digest under that assumption.