NODEDC_MISSION_CORE/docs/audits/2026-09-11-planning-project-browser.md

LAB planning project browser — 2026-09-11

Owner-approved job and composition

The LAB planner compares point-cloud passages before any potential mission-planning use. Its two entry paths are opening an existing experiment and creating a new project through +. The owner explicitly requested a large scene with an overlaid, movable/resizable inspector. No primary navigation or shared visual entity was added.

The former fixed editing column and separate comparison modal split one operator job across draft and result selectors. They are replaced with one header catalog Совмещённые маршруты, create, refresh and settings actions. A saved run opens its own immutable comparison immediately. New-project name, source, bounds, direction and query acquisition are inside the inspector. Settings open by default for creation/preparation; saved results open in review mode with settings available on demand. Closing settings clears the header action's pressed state without changing the selected scene.

Canonical Design Guideline exports: ApplicationPanel headerTools, Select, IconButton, WorkspaceWindow, Inspector/InspectorSelectField, TextField, RangeControl, SegmentedControl and LoadingRegion. WorkspaceWindow is the bounded scene tool admitted by WINDOWS_AND_LAYERS.md: it owns pointer/keyboard drag, resize, maximize and reclamping. Application CSS only arranges the domain content. The previous comparison Window component is removed.

Project identity and evidence

missions/projects.py projects existing stores; it creates no second project database. Each recorded registration or physical live experiment retains a separate kind:run_id identity, timestamp, frozen draft revision, sources, outcome and available evidence. Unstarted legacy drafts remain identifiable as preparation. Preparation-only cancelled live probes without a query are not presented as passage experiments.

Catalog and detail GETs never select an active live test, start acquisition, change a draft or rerun registration. Recorded scenes are hash-checked against the persisted report on every open. Large per-point correspondence arrays are omitted from the UI detail. Live historical views use the exact committed step's hash-bound input, source sequence, query path and fitted transform, with a separate derived-view cache. The terminal unregistered preview is never substituted for a saved fit. Missing result geometry remains unavailable.

The new project form requires a name and saved reference. Новый проход uses the existing planning-profile connection workflow. Из записи selects an already captured repeat and interval in the same inspector. Starting saves the setup, freezes it into a new run and selects that run. Viewing an existing finished project is read-only. Selection persists as optional session UI state; report authority stays on the server. Changing selection does not reorder the catalog.

Renderer correctness and retained limitations

The saved renderer applies T_reference_query once to both query cloud and query trajectory; the reference is unchanged. A synthetic nonzero translation test verifies this convention. Existing source clouds, transforms, thresholds and result reports were not modified or recomputed during this UI task. Double surfaces visible after registration remain evidence of geometric residuals or source-map effects; the UI does not hide them or claim they were eliminated.

The real retained B result is 820571ba-6076-482b-be34-29ceb5328c80, JA-SADOVAYA-001 → JA-SADOVAYA-002, 2026-09-11 17:24:04 Moscow. Its report SHA-256 remains 97236f89d73f63745a77558bf585dcac5ead5ffaa8191c27afb73b52c6271d59. The older failed live run remains a separate Без результата совмещения entry. Its stopped-time message is shown as historical context, not a fresh instruction to rescan. Prior all-green query rendering is explicitly labelled as the early format rather than per-point acceptance.

Code ownership

• Core/backend: missions/projects.py, web/mission_registration_api.py and router composition. Read-only catalog/detail and verified scene delivery.
• UI core: planningProjects.ts, usePlanningProjects.ts, useMissionPlanner.ts, useRegistrationTest.ts, PlanningTestContext.tsx.
• Domain components: PlanningProjectSettings.tsx, PlanningProjectResult.tsx, existing RegistrationScene.tsx. Removed MissionRegistrationWindow.tsx and its redundant UI polling/history path.
• Workspace: MissionPlannerWorkspace.tsx, feature CSS, typed header portal host through App/WorkspaceRenderer. API and project behavior remain outside App.

Verification

• 30 focused Python tests passed: frozen-project identity, separate run entries, preparation and failed-live states, scene digest rejection, read-only archive, committed-live-step rendering/cache, one-time query transform plus registration/planner/live regressions.
• 850 frontend tests passed. After final catalog-order/polling cleanup, 12 focused architecture/planner tests passed again. Full TypeScript check and production build passed; final build includes the small historical-message wording adjustment.
• Actual canonical 8000 browser: automatic saved-result opening, exact project selection, old recorded result, failed live entry, + with blank name and mandatory reference, query selection from saved A/B, launch disabled until valid fields, unsaved-form confirmation, settings close/reopen, keyboard movement and resizing, maximize/restore/Escape, ordinary/expanded panel, selection across page reload.
• Geometry check: inspector moved 20 px and resized from 390×600 to 400×610; stage remained 628×839.906 px. One Rerun iframe stayed mounted during settings interaction. Pointer drag/resize uses the existing canonical WorkspaceWindow; the new integration's measured geometry check used its keyboard controls.
• The temporary UI form was discarded. No new real registration, capture or device command was submitted. Three recorded reports and two live reports remained in the private stores; five visible projects were projected. No independent localization/false-positive/live-camera acceptance is implied.
• Canonical LaunchAgent was restarted to load the additive backend routes. A single backend remains on 127.0.0.1:8000; no listener on 8765. Docker backend/VM was not started. Tests/builds and browser QA ran sequentially; numerical workers were not launched for this UI change.

Next stage

Use the retained A/B evidence to test successive bounded windows and wrong-region controls. The candidate remains geometric agreement, not proven robot-pose accuracy. Hardware capture, obstacle policy and vehicle control remain separate acceptance work.