Complete route search before provisional localization

Decision and scope

Owner-approved follow-up to the entry/seam audit. Fix cold-start search scheduling, retain the precise start-area solver and all geometric/fresh-data gates. No device commands, threshold relaxation, map correction changes, endpoint snapping or retroactive edit of the failed study.

route-relocalization/v7 compares the complete selected route before choosing a provisional place. A finite hypothesis queue replaces the shared 35-second deadline. stationary-fresh-bootstrap/v4 separates search geometry from current localization evidence; an old prefix can never establish tracking directly.

Algorithm and boundaries

Collect the existing 10-second stationary prefix.
Run all 108 established dense-start seeds at the start/last confirmed hint. Keep a qualified dense candidate, but do not accept it before route comparison. A missing local start target does not prohibit searching the remaining map.
Index the complete reference and rank all geometrically usable route anchors, spaced by the existing 5-m policy. Ranking changes order, never eligibility.
At each anchor, test the existing three yaw proposals with two translations: sensor entry to route anchor, and the former cloud-median placement. One prepared local target tree is shared by all six fits. No atlas density cap or new thinning is introduced.
Compare the resulting SE(3) clusters together with the dense-start candidate. A distinct similarly good place remains ambiguous; first passing fit is not accepted. A failed fresh trial can advance to the next unambiguous candidate, with a new receipt fence and no reused validation data.
During search, continue checking source identity/order, receipt segment, pose/cloud freshness and stationary translation (existing 0.10-m envelope). Movement or broken input invalidates the hypothesis. The existing 40-second prefix-age fence remains for the legacy local protocol, not for v7's explicitly monitored stationary whole-route search.
After the result arrives, discard all pre-result receipts for confirmation. Three consistent fresh geometric checks remain mandatory before tracking. Ordinary tracking, local reference extraction, quality thresholds and the 8-second fresh-result gate are unchanged.

The numerical child emits progress after individual work units. The supervisor allows arbitrary total search duration while work advances; 60 seconds with no progress is a worker stall, not evidence that the location is unknown. STOP, source end, moved-search cancellation and teardown terminate/reap that exact child. A queued operator retry waits for its release and starts with a new prefix. Recording remains owned by the device plugin.

The candidate queue stays in one worker across scheduling batches; this is not a durable checkpoint/resume facility after a process crash. A failed worker returns incomplete search and offers reinitialization without stopping recording. Operator copy uses the existing status surface under the product UI canon.

Real saved-input regressions

Same immutable run 473870e0-5210-452c-b9e4-9d7937e93646, query 20260921T103511Z_viewer_live, corrected full reference JA-STROITEL-SUN-RING · correction v2, 579.98 m / 694,420 points. Only original initialization arrays enter the production worker. No end pose, later successful transform, manually selected anchor or offline deadline override. One CPU worker at a time; numerical thread counts fixed to one.

Original attempt	New result	Surface overlap	Inlier RMSE	Worker wall
First, failed cold start near route end	Candidate at anchor 113 / 565 m	98.8467%	0.14330 m	138.58 s
Second, familiar start	Candidate at start	98.7191%	0.13466 m	156.52 s

Both searches completed 810/810 fits: 108 dense + 117 anchors × 3 yaws × 2 seeds. No remaining route anchors; no competing qualified location close enough to trigger ambiguity. The second case retained the dense-start solution rather than degrading its target geometry.

All 159 files of the original saved study were hashed before/after each replay and remained unchanged. Raw evidence and generated arrays/results remain in the ignored private 2026-09-21-full-route-fix audit directory. The earlier study still truthfully records its failed first initialization.

Validation and review

121 focused backend tests passed: route coverage, last-ranked candidate, dense/global competition, median-seed regression, anchor-origin placement, multi-minute search, stationary continuity, distinct fresh windows, stale and ambiguous rejection, live recovery, cancellation, same-recording retry, child stall/forced cleanup and corrected-map source admission.
Focused Ruff and git diff --check passed.
Replacing median placement outright failed the existing lateral-offset GICP regression; therefore both initializations are retained. No quality threshold was weakened to make that test pass.
Registration/correction producer files and reference generations are unchanged.
Application architecture checks, TypeScript, all 887 frontend tests and the production build passed sequentially. Only the existing large-chunk build warning remains. Normal/expanded planner, latest ring evidence and Escape menu dismissal were checked in the existing in-app browser; no live state was faked.
With acquisition idle and the old study completed, the canonical LaunchAgent was restarted. Health/liveness passed on the sole port-8000 backend (PID 64086); no backend on 8765 and no temporary search/test workers remain. New captures use v7; the old report continues to expose its original v6 evidence.

Limitations and next field acceptance

Completeness is relative to the declared finite anchor/yaw/seed policy, not a mathematical proof that every pose in continuous space was searched. Descriptor ranking remains imperfect; queue completeness prevents a low rank from excluding the real place, but is not a fast kilometer-scale place-recognition index.

Full comparison increases even familiar-start latency to roughly 2–3 minutes on this 580-m map. This is an explicit regression in speed, accepted here for complete comparison, not a claim of final optimization. Target preprocessing is reused; safe coarse-to-fine acceleration still needs independent evidence.

Offline replay proves the corrected numerical cold-start outcome, not live stationary continuity or independent absolute pose accuracy. Synthetic tests prove the temporal contract; the operator must confirm it on a new capture. Surface overlap/RMSE are not position accuracy or autonomous-driving acceptance.

Field check: start at the formerly rejected place (last point of the old pass), leave the scanner stationary through search and fresh confirmation, then walk 20–30 m across the seam and back. Record the time to tracking and any interruption. No longer route is needed to answer this regression question.

7.0 KiB Raw Blame History Unescape Escape