# Mission planner: recorded zone and route draft

Date: 2026-09-11. Ops: MISSIONCOR-81. Canonical service: `http://127.0.0.1:8000`.

## Delivered boundary

The existing Missions/Planner workspace now selects a source from the shared observation-session catalog, reads its complete scanner trajectory, selects a contiguous interval in travelled metres and either direction, and persists a versioned draft without a vehicle. The original disabled five-step mock and its placeholder readiness badge are removed. The three blocks are Apparatus, Zone and Route. Parameters stay inside their blocks for this increment; the wider three-panel inspector composition remains a later refinement.

The cloud view reuses the existing bounded Rerun overview, top/3D and visual height clip. The route view uses an equal-axis XY schematic and a position slider over the selected original pose indices. This is inspection of a recorded scanner path, not measured localization or a certified chassis path. Selected interval and direction are editable; arbitrary waypoints, direct Rerun drawing and timed playback are not implemented.

## Source and persistence contracts

- `missions/sources.py`: immutable generation from validated source identity, approved-prefix staging, input digests before/after preparation and source-bound cache. One preparation at a time, maximum 100,000 poses / 30 MiB derived JSON, parser time/message bounds.
- The plugin's `planning_source.py` preserves all pose indices, message ordinals, metric positions, elapsed arrival times when available, and cumulative distance. No second pose transform is applied. One decoded nonempty point-cloud frame confirms spatial evidence. Other cloud messages are counted without decoding every point again; this is not a full point-cloud integrity/accuracy audit. Full overview diagnostics remain separate.
- `missions/drafts.py`: SQLite transactions and compare-and-swap revisions, frozen source identity, selected source indices and resolved points. Changing/missing source never removes a saved draft.
- `web/mission_planner_api.py`: bounded request models; no vehicle ID or execution parameter is accepted. Data checks bind an exact saved revision, reverify source digests, measure length and largest position step, and retain an immutable report. A concurrent draft change rejects publication of a stale report.
- `core/missions`, `components/missions`, `workspaces/missions`: separated domain contracts, reusable renderers and workspace composition. Existing Design Guideline Select, RangeControl, TextField, Window and actions are reused.
- Raw captures, cached poses, SQLite drafts/checks and real scene media remain runtime data outside Git. Nothing is installed on the board.

## Validation

- 846 frontend tests passed before final bounded UX refinements. The subsequent metre-input change passed the five focused planner tests plus four architecture tests. Final production build includes TypeScript checking.
- 16 focused backend tests passed (planner + existing overview); 10 planner tests passed again after removing unnecessary cloud decoding.
- First real source preparation after the optimization: HTTP 200 in 2.35 seconds for 4,959 poses and a 487.61 m recorded path. This is a local functional measurement, not a board or localization benchmark.
- Browser on the actual 8000: selected the existing road session, obtained the initial 30.01 m / 578-pose interval, saved a named draft without a vehicle, ran its data check (matching source digests; largest step about 0.10 m), reloaded the client and reopened it. Reverse direction starts at the selected final source pose. Exact interval entry uses metres, not hidden frame indices.
- Another source named `35` loaded independently with 283 poses / 0.089 m, proving the selector is not bound to the road session. Unsupported derived LAB entries remain visible with a reason.
- UI QA used the actual narrow in-app browser surface and expanded window, including Escape. Desktop-width rendering is not separately qualified here.
- All checks/builds/browser work ran sequentially. No duplicate backend or port 8765 was introduced. Canonical 8000 remains durable operator state.

## Next experiment

No new scan is required to inspect the saved draft. Before claiming live testing, implement a bounded geometric-registration runner, prove known synthetic transforms and reject a wrong reference segment. Then acquire an independent B recording in a new K1 project over the same 20–30 m section, retaining both Mission Core session and native capture. Freeze reference A. Compare B windows against A without future B frames, report failures/ambiguity, timing and memory. Only after offline evidence passes should the same pipeline feed a live testing viewport.

The current “Проверить маршрут” action does not compare clouds, estimate the scanner's location, determine passability, detect obstacles or send commands. The starting-position tolerance and target-board capacity remain experimental questions.