nx_editstack_lib.nx

nx_editstack_lib.nx -- THE NON-DESTRUCTIVE OPERATION STACK: record, undo, redo, replay, re-evaluate. This is the spine every content-creation tool is built on (Blender's modifier stack, ZBrush's layers, Houdini's graph) and the one thing this estate has never had. /compare/dcc DC1 and DC2. CHECK-BEFORE-BUILD, and the incumbent was READ, not guessed. nx_spendgate named nx_editor_canvas.nx as the nearest candidate and it is NOT this: it is a 2D page-layout canvas ([kind,x,y,w,h] elements -- heading, text, button, image) whose undo is ec_save/ec_undo, a ONE-DEEP FULL-ARRAY SNAPSHOT copied in and out. That cannot redo, cannot replay, cannot re-evaluate a parameter at depth, and costs O(state) per step. nx_siteedit_history.nx is file REVISION history, a different layer again. So the estate's only existing undo is one-deep and snapshot-shaped; this is the log-shaped primitive underneath, and nx_editor_canvas is a candidate consumer of it rather than a rival to it. WHY UNDO IS REPLAY-FROM-BASE AND NOT AN INVERSE OP. Applying an inverse operation is how a float editor loses precision: scale by 3/7 then by 7/3 does not return you to where you started, so undo DRIFTS and the drift is invisible because nobody digests the state. Replaying the log from the base cannot drift, by construction, and it is what makes the bit-exact claim on the /compare/dcc board checkable rather than asserted. The cost is O(cells x ops) per undo, paid deliberately and named here so the next reader does not mistake it for an oversight. THE STATE IS AN i64 CELL VECTOR ON PURPOSE. A consumer maps its own data onto cells (a mesh maps vertex components, the canvas above would map [kind,x,y,w,h]), so there is ONE stack in the estate rather than one per subject. All integer: no float enters, which is why replay is bit-exact. 100% sovereign. No hardware writes (Rule 26). license_tier: ORIGINAL