DCCONSTRUCTIONS
e515ab1b8c
Preserve the completed teach-and-repeat laboratory stage: reference preparation, cascaded acquisition, local tracking and recovery, recording lifecycle, replay qualification, and persistent Rerun scene controls. Document the open grid-picking regression and Rerun upgrade contract. No autonomous driving or loop-closure optimization is claimed.
5.1 KiB
Planning scene and independent pass B — repair audit
User scope: retain the ordinary spatial scene toolbar, camera, recording controls and Planning profile; eliminate the exit action that loses the research; unify the planner toolbar and Inspector geometry. No new capture or hardware command.
Retained evidence
JA-SADOVAYA-002 was captured on 2026-09-11, 13:43:55–13:47:03 UTC. The archive
catalog marks its point-cloud and trajectory sources ready: 959 cloud messages,
960 poses, 37,812,938 raw bytes. Raw SHA-256:
fe1366a177c20204f26e709649521611f0fc3aae138cb857bcabe84a5809268a.
The device's last distance report is 62.532 m; sum of received pose steps is
62.974 m, including gaps, and is not independent ground truth.
The research failed before its first registration result: 2.776 m of accumulated movement and 14,101 preview points were retained in its status. Received pose 161 moves 10.407 m after a 9.973 s receipt gap. The first implementation rejected any step above 3 m without considering elapsed time. Therefore no green was produced; this is an ingestion failure, not an experiment disproving localization.
The selected frozen draft was the prior internal reverse-direction comparison of the first 30 m of A, not the forward 30 m draft. This source selection is preserved, not silently corrected. Start/direction must be checked before using B to claim a registration result.
B has no archived camera frames. Its camera epoch sealed with zero media segments
and stale-activation-commit. This is separate from the missing floating camera
UI. Camera authority fences are retained; no unsafe retry or physical command is
introduced by this patch. An unavailable recording cannot be reconstructed as
video. The next actual camera run still requires hardware acceptance.
Implementation
SpatialWorkspace is extracted unchanged in ownership from the workspace hub to
workspaces/spatial. A typed visual-profile slot supplies the research renderer,
status, metrics and toolbar; the existing K1 controls, source picker and floating
camera transport remain owned by the shared scene. The profile no longer puts
stop controls behind an extra modal or exposes a button clearing the research.
Terminal errors outrank generic waiting/staleness. Saved research selection
restores its frozen reference and verified preview across navigation/restart,
without running a new capture or fit. Future previews are persisted with hashes.
Receipt gaps above 2 s clear the bounded fit window and invalidate old accepted samples. A step is rejected above max(3 m, 3 m/s × min(receipt gap, 30 s)); this is a diagnostic plausibility bound, not odometry acceptance. Jobs carry the segment identity and cannot commit across a gap. Slow results remain recorded as historical evidence; the existing 8 s green freshness rule is unchanged.
The planner uses canonical Inspector/InspectorSelectField with full-width stacked actions. One right-aligned host toolbar owns cloud/route, top/3D and expand actions; no nested host title. For the pinned single-view Rerun runtime, its measured 54 CSS px canvas chrome (28 px top allocation + 26 px view strip) is allocated outside the clipped content viewport. The Inspector scrolls independently inside the available panel height. Scene content height and native pointer handling are preserved.
Validation
Focused Python registration/planner/live tests pass (24 cases), including the 10 s receipt-gap regression, rejection of a fast coordinate jump, bounded sample reset, persisted preview restoration and no restored live-green authority. Frontend architecture and planner checks pass. Full frontend validation and canonical browser smoke are recorded at completion below.
The original failing report and raw session are retained. No localization success, new camera capture, autonomy, control or navigation acceptance is claimed.
Completed acceptance on canonical 8000
Production typecheck/build passed. The complete frontend run executed 849 cases (840 passed; nine obsolete architecture/source-location expectations failed). After updating those expectations for LAB placement/shared scene extraction and removing generic-shell vendor labels, focused reruns passed; the final remaining four affected files passed all 50 cases. No failing assertions remain from that run.
Browser inspection on the built localhost service verified: equal action widths (315.2 CSS px in the observed Inspector), split canonical selects, independent Inspector scrolling, one aligned toolbar, no visible Rerun title/help strip, real A cloud, expand/restore/Escape, preserved forward draft, and reopening the actual failed research with its frozen reference and explicit failure. The old failed run has no persisted query preview; its 14,101 status points are not reconstructed or presented as a successful result. Raw B remains intact.
The shared source picker shows K1 cloud and both camera channels; cameras are unconfirmed/disabled while capture is stopped. Camera/recording controls retain the ordinary scene owners, but a new physical camera run was not performed. The next action is analysis of existing A/B in the correct forward draft, not an immediate repeat scan. This UI repair is not hardware-camera acceptance.