11 KiB
Explicit Allocation Arrange amendment — 2026-09-23
Status: accepted design; pure planner verified, App integration and native acceptance pending. Codex-9e8r6t /root adopts this bounded policy after reviewing the actual OA/SA experiment, its independent critique and corrected geometry oracles. No user-owned application specification is changed.
User-derived basis: PROJECT_GOAL, multiselect contract. Existing authority: Allocation drag v2 and its linked destination-feedback amendment. Experiment and planner evidence records the decision history, exact source, results and limits.
Scope of the change
Explicit Arrange only supports one or more selected work/performer placements through one declared footprint and one combined geometry check. A distributed work root does not fail Arrange solely because its allocation summary is MixedOrUnresolved or its semantic descendants occupy different visual performer groups. Its affected projected footprint must still translate rigidly and satisfy all geometry/readability checks below.
This narrowly supersedes v2's mixed-composite/coherence pre-veto only for explicit Arrange. Ordinary work-body assignment is unchanged: work-only groups use one frozen common destination, normalize work ancestry, and admit each root under existing role/coherence rules. A mixed/distributed composite still refuses ordinary assignment; different valid initial performers on independent roots do not alone require refusal. Same-effective-performer role-only roots gain no direct Allocation, unsupported role exits refuse wholly, and no drag changes semantic work ownership. Ordinary mixed work/performer selection cannot partially assign its eligible members. Offer Arrange only when its own full-set admission is available.
Arrange admission cannot depend on having two selected items. C alone and {A,C} use the same rule; adding A changes the actual movement footprint rather than unlocking an otherwise identical operation. Existing ordinary performer arrangement remains governed by its existing behavior; this amendment defines the explicit mode consistently across cardinalities.
Exact target and footprint definitions
- Validate every original typed target before normalization: exact placement/lifetime in the same active Allocation diagram and layer, currently visible, eligible work or performer kind, and core-valid accepted Project. Any stale, missing, foreign, port, edge or grouping-relation member refuses the whole action. Same-element duplicate placement in one diagram is malformed input. Preserve exact selected membership and primary separately from planning roots.
- Normalize explicitly selected work roots by same-kind semantic ancestry. For each root, S contains its recursive same-kind work closure. Q is all existing active-diagram placements of the union S, including folded descendants. Unplaced work contributes no fabricated placement.
- V contains each explicitly selected performer placement and its accepted visual descendants, including folded descendants. R = Q ∪ V. Capture this union once; do not acquire semantic ancestors/siblings or new semantic roots merely because a work placement is an incidental visual descendant. Each R placement moves once. Authored placements outside R and every other diagram remain exact.
- The press anchor must be an exact selected placement. Snap its requested position once on the existing ten-world-unit grid and apply that one delta to all R. Another selected anchor may change the snapped delta but never the mode or R. Preserve authored sizes, fold states and port anchors. Arrange authorizes no destination unfold, reparenting, packing, Fit or automatic topology change.
- Preserve every engineering record, including Allocations, RoleAssignments, RoleRealizations and semantic parents. Resolve routes using accepted actual endpoint owners and the complete R delta map: internal pins with both owners moving by the same nonzero delta translate once; one-ended pins remain fixed. Relative labels and attachments remain exact.
Combined geometry and its deliberate diagnostic boundary
Project the complete accepted and candidate diagrams using the appropriate semantic index. Check the entire resulting visible scene, not only moved cards. An unselected performer enclosure may change its derived bounds even though its authored record is untouched. Such growth is allowed only if the complete candidate meets model bounds and the existing unrelated-overlap rule: no new or enlarged positive-area overlap beyond the current 0.01 area tolerance; visual ancestor/descendant containment is exempt.
Also compare accepted/candidate diagnostic scenes. In each scene, temporarily unfold R and the visual ancestors necessary to reveal R; use the corresponding accepted/candidate visual ancestry. Include stationary geometry revealed by that required expansion. These temporary Diagram changes are never committed. A moved hidden child cannot evade collision checking through a selected performer's fold.
Unrelated stationary folded interiors are outside this diagnostic domain. Do not unfold every container in the diagram. Their visible collapsed headers still participate in the ordinary scene check. This limit prevents invisible distant contents from becoming an unexplained blocker; it does not promise that future expansion will be free of overlap. A stricter all-unfolded policy is a different candidate, not an implicit interpretation of this amendment.
Every R node's affected diagnostic footprint must preserve size and translate by the common delta. No automatic convergence/reflow is licensed by removing the semantic pre-veto. Check the required originally selected headers for visibility and containment in the captured viewport; a normalized-away selected descendant cannot silently disappear. Any unsupported footprint, unreadable selection, invalid bounds or new/enlarged unrelated overlap refuses the complete action. Show the actual combined candidate and affected counts; do not present independently valid singleton ghosts as evidence of group feasibility.
Valid and refused examples
| Example | Required result under this amendment |
|---|---|
| Composite C contains x on A and y on B; select C; enough room for one rigid delta | Arrange may move Q={C,x,y}, preserving A/B authored positions and every allocation. Their derived enclosures may grow if the complete scenes remain admissible. Ordinary assignment still refuses the distributed root. |
| Same model, select {A,C} | R additionally includes A and its accepted visual descendants. Move each union member once. The operation is not equivalent to C alone. |
| Select parent C and its already covered child x | Validate both selected targets, deduplicate planning scope, move each placement once, retain the two exact selection identities/primary. |
| X's translated bounds enter Y's old bounds, while both selected roots move clear together | Check simultaneous candidate geometry. Do not refuse because an independent singleton-X check treated Y as stationary. |
| Unselected B's projected enclosure grows into stationary Z although no R card intersects Z | Refuse. Unchanged authored B geometry is not evidence that its derived enclosure is harmless. |
| Folded selected A hides moved x, which newly intersects a stationary visible obstacle | Required diagnostic expansion detects the collision and refuses. |
| Unrelated folded D hides Z that intersects moved x only in an all-unfolded alternative | Z alone does not block under this diagnostic policy. Required visible/R-plus-ancestor checks must still pass; retain D/Z authored records and D's fold. Future expansion may reveal overlap. |
| Selection includes a work card plus port, edge or grouping-relation target | Whole-set refusal; no eligible-subset fallback. |
| Snapped zero delta, unchanged semantics/folds/routes | No command and no history entry. Arrange never changes assignment merely because a pointer crosses a performer. |
Lifecycle, source and history
Retain v2's clean-owner, dirty-pointer shielding, modal/tool priority and frozen camera/input lifetime. ArrangeNext is explicit and consumed once at press, whether the anchor is work or performer. Preserve the selected set and primary on selected-member press while distinguishing stationary release from group drag; a drag cannot become a singleton click after crossing the three-screen-pixel threshold and returning near its origin. Mode and footprint cannot change midway through a token.
Capture exact selection generation with document/revision, active diagram/layer, epoch, bounds/pan/zoom and accepted scope. Dirty source/property/behavior input, owner/tool/lifetime change, Fit/resize, Escape, focus loss or outside/hidden release cancels without partial movement. Retiring a token is irreversible even when numeric values later return. Unsupported mixtures never fall through to legacy MoveNode, assignment, reconnection or deletion.
Preview is immutable. On release, revalidate the same capability/snapshot, recompute combined geometry and compare the complete prepared Project with an independently constructed expected Project containing only the permitted R/route edits. Execute one Batch through normal Document source admission. Source failure/refusal leaves accepted source, selection/primary and history unchanged; successful movement is one Undo/Redo step with exact source roundtrip. Geometry detector assertions alone do not establish any of these transaction/lifecycle properties.
Choice rationale and rejected alternatives
Retaining strict v2 is the smallest implementation but rejects a semantics-preserving explicit layout solely for distribution; until this amendment is implemented, its refusal remains truthful. Visual-selected-only translation is rejected because it omits required semantic/folded descendants. Sequential singleton planning is rejected because geometry, routes, deduplication and final expected Projects are not compositional. Universal unfolding is deliberately not chosen for this bounded mode; it introduces invisible unrelated geometry into admission. Packing/reflow and arbitrary convergence remain separate layout work.
Evidence and acceptance still required
The experiment/planner report is authoritative for exact runs, source identities, independent review and retained failures. It distinguishes actual projection/source/history evidence from policy predictions and App acceptance. Its independent expected geometry accounts for SA function ports; collision detection is not presented as atomic UI refusal.
Production integration still requires:
- Whole original typed-set admission, token/selection-generation, dirty/stale/source-failure, one-shot mode, irreversible threshold, cancellation and minimum-readable-text-scale checks.
- OA/SA native workflows anchored on performer and work members, truthful counts/preview/refusal, one Undo/Redo, Save/reopen and exact unrelated-record preservation.
- Separate ordinary work-only group assignment with per-root responsibility rules and one combined candidate/geometry gate.
This amendment defines a bounded Arrange capability. It does not establish all Allocation group interactions, general convergence, grouping removal, routing usability or completion of the multiselect issue.