ARCADIA-Studio/harness/release-final-native-priorities-20260906.md

Final native priorities — 2026-09-06

Read-only triage against the objective, current STATE/TASK_LEDGER, coverage audit, and later native reports. No Cargo or native operation. These are eight bounded actions, not a Cartesian product of layer, platform, size and every enum. Root should use the newly accepted binary/assets and retain its exact hash. The Linux worker's current application gate proceeds independently.

Prioritized acceptance matrix

Priority / classification	Existing fixture and smallest useful task	Required observation / stop condition
1 — known material defect, source fix awaiting native acceptance	fixtures/release1-native/communication-means/cm-reverse-scenario.arcadia; repeat only the reverse scenario transition on the new binary, using its existing request/picker instructions and distinct check reverse-behavior mode.	Apply selects the explicit OA scenario counterpart in Behavior, preserves Call/Reply correlation and all saved diagrams, then exact Save/Undo/Redo. New oracle expects ten generated target/trace UUIDs and zero placements. Preserve legacy reverse's passed fifteen-UUID/five-placement result. No repeat of drawn CM/create-one/reuse-both trials.
2 — new implementation and corrected teaching assets, untested together	Copy complete examples/release1/; exercise Home/guide Open walkthrough at1100×700, then open the working copy. Follow README sections1,4–6: responsibility versus work, mission/include context, interface/data contract, deliberate warning, and reviewed oa_confirm→sa_confirm name edit.	Dirty Cancel retains raw input; opening/replacing uses the normal guard. All30 source modules and evidence target remain available; original packaged folder unchanged. Exactly the located warning in expected-diagnostics.json; inspect/refine it through its located object. Reviewed rename updates only intended source/history; Save/reopen preserves it. Guide actions remain reachable. This is the practical onboarding/parity story, not another full CM replay.
3 — implemented behavior editing; known compact-form correction lacks native closure	Working copy of examples/release1/project.arcadia, Process oa_response and referenced oa_field_procedure, plus Scenario oa_full_scenario/oa_conversation. At1100×700 open a Two paths/contribution form, edit both path values/guards, Cancel then commit one bounded change. Inspect and edit one explicit reuse binding. Use fixtures/release1-native/semantic-ordered-fragment.arcadia only for the still-uncovered Fragment branches final-order control.	Form fields scroll independently; Apply/Cancel stay visible; canvas/owner/camera remain stable. Real pointer gesture produces correct same-context links and no duplicated shared work; rejected reuse/order retains drafts. One exact model/source/history comparison per distinct committed change. Fragment order uses its supplied ambiguous/final request and .after-order oracle; source5→6 header migration must be explicit. No need to repeat main/reverse/Scenario ordered checks or every ALT/PAR/LOOP combination.
4 — known inspector correction plus untested saved-category lifecycle	fixtures/release1-native/dense-40/dense-40.arcadia or the retained .local/native-qa/source6-dense-result.arcadia original10/copy9 result. Switch saved original/copy while inspecting the view; make a category membership edit and delete that category on a disposable copy.	Inspector follows the selected view; an unfinished draft obeys Apply/rejection guards. Category deletion keeps semantic exchanges and removes/repairs references through the documented command. Undo restores category/filter membership, placement/route facts stay exact, Save/reopen preserves the result. Use dense-40-oracle.json for existing visibility cases; derive a specific command oracle before the new lifecycle edit. Do not repeat forty-label inspection or all eleven filter matrices.
5 — bounded state presentation gap, not absent typed actions	fixtures/release1-native/state-actions.arcadia: fold/unfold the nested region, rename one linked work target, inspect its typed action row, Save/reopen the saved state view.	Typed action retains its real relation identity and refreshed name; Internal transition/nested region coexist; saved view routes back to Behavior. Exact source/history for the rename; folding preserves semantics. All three Entry/Do/Exit selections, compact Exit selection, timed-event rejection/recovery and direct state-create/Undo already have native evidence—do not repeat those whole trials.
6 — corrected artifact wording only	fixtures/release1-native/artifacts/semantic-artifacts-create.arcadia: reserve/stage the explicit destination; Take Presence plus Keep Owner to expose the reviewed conflict, then repair or Cancel. If needed, use semantic-artifacts-move.arcadia for kept Presence versus moved Owner.	Reserved alias/name is readable in picker, facet title and conflict; no false Deleted reference label. Same-row facets are identified accurately, Apply disabled for conflict and no model mutation. This short display regression closes the new wording; CREATE/MOVE/DELETE exact history and raw crash recovery already passed and need no replay.
7 — fixed hosted destination isolation and current source6 integration	Publish a disposable complete examples/release1 SourceBundle to a fresh loopback store. From its hosted session open a copied entry with a deliberate syntax error, repair it and Save locally; separately cancel/fail a local Open.	Repair-workspace Save never changes the old server resource/ETag; Cancel/read failure retains the hosted destination. Catalog clearly reports stale content after a write and explicit Refresh updates it. The lesson's source6 categories/CM/artifact fields survive server import/read/restart. Existing real stale-ETag conflict, offline retry and local-copy native trials already passed; repeat them only if this targeted run exposes a new transport regression. No old D:/ fixture is required.
8 — platform acceptance, not a Windows rerun of every feature	On the current Linux application build, use the same complete lesson working copy for one compact direct behavior edit and one typed data/interface edit, then Save/reopen and normal close. Use its model.json/source bundle for the Linux server import/validation/restart gate already assigned to the worker.	Working renderer/input at supported size, readable pending forms, exact persisted model/source and no serious runtime output. Cover one still-unverified data association/port-allocation operation if its contract is selected before the trial; no invented expected output. Prior ext4 core/syscall passes and older Wayland native trials do not establish current GUI/server/package acceptance. A headless guest server pass alone cannot close desktop support.

What this matrix deliberately does not reopen

Later reports supersede the old coverage audit's missing-CM, missing-selective-view, missing-completeness and missing-semantic-facet statements. Those implementations exist. Native batch, Child/Incident, source5 recovery, artifact lifecycle, keyed Take/repaired Binding, CM drawing/port alternatives, ordered main/reverse/Scenario, guide navigation, dense duplicate/hide/reopen, and state timed-draft recovery are scoped passes. A new executable hash requires targeted changed-path checks and final automated regression, not automatic repetition of every historical task.

Optional permutations after these gates include every pseudostate/trigger, both directions of every scalar/attachment facet, each fragment operator at both sizes, every filter intersection, all role/capability graph permutations, and a new crash cycle for every already-tested raw adapter. Escalate only when a distinct uncovered control or observed failure gives such a case a concrete purpose. Branch-steps order shares the exercised ordered-row engine; Fragment branches is the higher value remaining native collection shape. Automated contract coverage remains necessary for both.

Scope reconciliation and non-native release gates

• docs/scope.md and the fixture root README still call artifact/attachment/CM native work pending, despite later passes. The historical audit and append-only native report also retain superseded gap text. Update current-facing summaries by pointing to the later exact reports; do not erase historical evidence.
• F2 is a root scope decision, not a user-approved reduction or a passed behavior-export feature. docs/scope.md and the guide still use broad “SVG export”/“diagram image” wording; explicitly qualify architecture-only export. The objective asks for fully feature-complete OA/SA according to the application specification, while scope.md describes a bounded Studio profile. Final domain review must resolve whether each excluded semantic capability is outside that specification; passing this matrix cannot settle that interpretation alone.
• Final current-tree full-workspace/MSRV/docs/OpenAPI/source/example checks, optimized Windows/Linux artifacts, notices/runtime/complete lesson packaging, clean-environment launches and fresh independent end-state review remain release gates. These should be scheduled once on the accepted tree, not added as another set of native semantic permutations. Existing QA packages and old aggregate test totals are not final distribution evidence.

Primary evidence: STATE, coverage baseline, current native, CM, artifacts, hosted, process, independent navigation review.