ARCADIA-Studio/harness/release-activity-allocation-drag-review-20260911.md

Independent allocation-drag proposal review — 2026-09-11

Codex-956atf /root/allocation_drag_critic, bounded source/design critique of the proposal for issue 3. Read local AGENTS/PROJECT_GOAL, remote main AGENTS, STATE, issue 3/comments and issue 4 coordination. Working branch read from .git/HEAD: codex/behavior-authoring-20260910, HEAD 06ec9621297627a1f7d5c3bc65efb237d1c9923e. Source anchors below refer to the files inspected in that workspace. Only this new report was written. No build, test, native execution, production edit, Git/index change or Gitea write occurred. This is not implementation or acceptance evidence. Root owns synthesis and integration.

Judgment

The diagnosis and principal choice are sound: ordinary drag in an Allocation view should compose allocation and placement. Existing MoveNode only moves placements/routes; projected parentage comes from allocation_summary and semantic ancestry. Work-to-performer Reparent, relaxed leaf-only validation, or turning off enclosure would damage the existing model. Direct allocation retargeting should preserve relation identity and metadata. Preserve separate decomposition, effective role responsibility and actual direct allocation.

The proposal needs the following corrections before it is precise enough to implement. The first three can otherwise cause hidden semantic changes or a broken ordinary gesture.

Material corrections

1. Resolve same-performer role semantics. A leaf with RoleAssignment R → x and RoleRealization A → R is displayed under A without direct Allocation. The gesture text promises Arrange within A, but the no-direct-allocation table says to create A → x on that drop. These are observably different: direct allocation subsequently overrides role changes, and OA communication allocation uses actual allocation rather than role responsibility. Specify that dropping within the current unique effective performer is layout-only for ordinary drag, including uniformly role-derived composites. Creating a direct assignment to that same performer requires an explicit Assign action. Dropping onto a different B may create direct allocation with the existing advisory role warning. If removing direct A reveals role-derived B, the label must say Remove direct assignment; role places work in B, not Unallocated. A role-only outward drop should explain its retained role grouping and leave accepted placement unchanged when it cannot fulfill the requested exit. Sources: effective performer, role conflict is a warning, actual OA communication allocation.

2. Compare surviving semantic records as well as deleted IDs. deletion_impact is not a complete semantic-change report. Counterexample: same-layer Constraint C references Allocation a in constrained. Deleting a clears C's reference while retaining C; the candidate can remain valid and the removed-ID set can be exactly {a}. The proposed extra-deletion gate would approve a silent constraint edit. Compare the prepared candidate's full semantic diff to the intended allocation edits, and expose/refuse any other semantic mutation as well as cascaded deletion. Presentation cleanup of the deleted relation's own route is a distinct expected effect. Full validation remains necessary for retargeting: state availability/actions and participant scenarios also depend on effective performer. Preserve allowed warnings; notably OA unallocation may leave a communication mapping incomplete with a warning, whereas SA can cascade through port and exchange allocations. Sources: constraint validity, surviving-record cleanup, deferred deletion, state context.

3. Guard mouse-down selection, not just release. Current node press arms state.drag and publishes SelectNode without checking editable. SelectNode can call apply() before a drop happens. Merely removing apply() from the new release handler does not prevent unrelated dirty properties from entering history, and may invalidate the captured revision immediately. Give semantic drag a clean-owner admission/selection path at press and recheck it at release. Dirty-source/property/behavior refusal must not fall through to legacy movement that applies the draft. Preserve simple selection behavior deliberately; do not silently commit a form to make drag work. Sources: press, selection applies, legacy release dispatch.

4. Specify two scopes and an actual placement rule for composites. Define the semantic same-kind work closure/leaf set separately from the presentation placements that move. Counterexample: mixed composite P has x allocated to A and y to B. x and y can be projected directly under those performers, outside P's visual subtree. Changing both allocations to C while moving only P's captured visual descendants leaves x/y at their old distant coordinates; C then encloses all of them. Even translating all semantic placements by one delta preserves that excessive spread. Specify a bounded packing rule for this converging case, or retain a truthful refusal until it exists. Capture hidden placements too; never invent placements for unplaced leaves, and edit only the active diagram's geometry. An outgoing child can also make its previously uniform composite move to a different projected group: include affected ancestor grouping in the preview. Existing insertion adjustment only pushes colliding siblings down; it does not solve arbitrary distributed subtree packing. Sources: parent derivation, descendant movement, insertion adjustment.

5. Keep the target contract stable and the candidate projection fresh. Press-time rectangles are the right defense against enclosure chasing. Keep their destination outline visible while candidate enclosures change, or the visible target and accepted hit area can disagree. Define semantic ancestor containment for nested performer precedence; geometric inclusion alone must not turn overlapping peers into parent/child. Define the boundary/overlap and work-card occlusion cases explicitly. Capture/invalidate bounds, camera and view lifetime as proposed, preserving the ordered native pointer handling. Semantic preview also requires a candidate Project and candidate ViewIndex: the current preview clones only Diagram and reuses the accepted semantic index. Extend scene-cache keys with semantic destination/plan identity, so an unchanged coordinate cannot reuse a previous target's scene. Recompute model work when destination/scope changes, rather than cloning and validating a large project on every pointer pixel. Sources: existing preview/cache, index lifetime contract.

Smallest coherent implementation path

Keep one pure, bounded proposal composer plus stamped canvas gesture/App admission. Use current commands and Document::execute; its source compilation precedes its single commit and history stores exact source snapshots (transaction). No metamodel change, allocation inheritance or new generic draft framework is needed.

First establish ordinary incoming, between-performer, outgoing and same-performer leaf gestures, then uniform composites with explicit descendant scope. Keep unsupported role exits, collateral semantic edits and unreadable mixed layouts as named local refusals with existing repair navigation. This is a staged scope, not grounds to claim every composite workflow complete. A full mixed-impact editor and automatic dependency repair are larger capabilities and should not block proving the user's basic direct workflow. Mixed status alone is not a semantic error or a reason for obligatory confirmation once a safe, readable exact-set preview exists. Extra Apply steps should resolve a real scope/impact decision. A visible Arrange only alternative should be temporary or clearly indicated, not a sticky mode users must remember after every assignment.

Acceptance additions before claiming the fix

• OA and SA leaf incoming/reassign/outgoing: exact relation identities and metadata, work parents, active-diagram positions/routes, save/reopen, and one exact Project/source Undo/Redo step. A snapped no-op creates no history.
• Role-only same-performer arrangement creates no direct allocation. Direct removal that reveals another role performer has the same grouping and wording in preview and result; warnings remain advisory.
• Constraint referencing an allocation, artifact/trace attached to an allocation, SA port/exchange dependencies, OA communication mapping and participant-bound scenario/state behavior: no unshown semantic deletion or surviving-record edit.
• Dirty form before press, dirty source, modal/save ownership, revision change, peer first-child allocation/role-assignment transfer, camera/bounds change, Escape/unfocus and outside release: no hidden apply or partial history. Include the actual mounted press → SelectNode → move → release dispatch, not only planner calls.
• Uniform folded/unplaced children and mixed P/x/y fixture: exact leaf scope, hidden-placement handling, bounded readable grouping or explicit refusal, unchanged other-diagram geometry, and no target/enclosure chasing. Verify competing/overlapping/nested targets and that the shown destination is the committed destination.

Exercise the completed direct workflow in the actual Windows package at the minimum client size. Planner tests and this source review cannot establish target discoverability, cognitive load, readable layout or native event behavior.