NODEDC_MISSION_CORE/docs/audits/2026-09-21-shared-ring-closure.md

Shared known-revisit correction for small and large rings

Scope and topology correction

The operator clarified that the small ring is inside the large ring and shares roughly half its path. Their unshared corridors must not be scored as a failed repeat of the same map. Reference surveys in Data and independent captures owned by planner studies are distinct evidence classes.

The earlier 30/101 replay used the planner-owned JA-STROITEL-RING-002 full small-ring pass, not the Data reference itself. Its whole path does not become covered merely because its origin is the planner. The earlier 29/29 short pass also came from the planner. Preserve the negative numbers, but classify their geometric scope correctly. Other short planner passes are separate queries.

Implementation and boundaries

reconstruction/closure.py owns a device-independent, injected-registrar policy for acquisition of a known start-area revisit. It has no hardware, UI, catalog or planner import. The K1-specific reconstruct_recorded_ring.py adapter now uses this same policy on both source recordings, with no source-name branch.

The separate offline first-fit registration policy admits a larger initial correction (25 m / 180 degrees) without modifying live tracking's 3 m / 30-degree policy or the existing overlap, RMSE, shape, information and correspondence criteria. These are declared acquisition bounds, not a promise of arbitrary drift recovery. Identity is the initial hypothesis; endpoint positions are never forced together or used to manufacture a seam transform.

All 12 training-only support pairs are evaluated: first [10,20,40] seconds × last [5,10,20,30] seconds. Held-out frames cannot enter either side; shared source frames are rejected. Each qualified forward fit must pass seeded reverse refinement and the existing 0.1 m / 0.2-degree round-trip check. Qualified support windows must also agree at both observed patch centers within the separately declared 0.5 m / 1-degree consistency envelope. This is an engineering ambiguity guard, not a calibrated accuracy claim.

The established 20/5-second measurement is retained if qualified. Otherwise the largest qualifying support is selected. All windows are still examined for contradiction before selection is accepted. The smooth field solver, rigid per-frame transform, original-frame clock binding and full-resolution export are unchanged. Failure keeps the audit and produces no silently accepted identity correction.

The adapter's v2 contract binds closure search and its declared policy. Sealed v1 decoded caches remain reusable only when all actual decode/holdout settings match. review_recorded_ring.py adds explicit frozen-data acceptance: all local windows qualify, enough seam holdout exists, fixed quality gates pass, and cross-visit agreement is not degraded. Packaging requires positive v2 acquisition and review, with the complete search included in the evidence seal. Previously published v1 map bundles and pinned studies remain supported.

This is an offline correction/admission workflow, not a newly installed background auto-corrector for every capture. It does not discover arbitrary interior loops or recover a measured drift history from native LiDAR sweeps.

Regression caught and retained

The first generalization always selected the largest qualifying support. It passed same-source checks (small 52/52, large 136/136), but warm replay of the independent full small-ring planner pass accepted only 100/101 snapshots. The 96.115 m window failed numerical convergence despite 99.866% overlap and 0.13993 m RMSE; the following window recovered. The same replay against its frozen previous reference accepted 101/101. This was a real numerical regression, not an uncovered-region excuse or permission to ignore convergence.

The final cascade preserves the qualifying established window and uses expanded support only when necessary. It reproduces the previous small-ring corrected points and trajectory bit-for-bit, and the previous successful large-ring diagnostic geometry bit-for-bit. The unsuccessful intermediate candidate and its full replay are retained privately; no quality threshold was weakened.

Verification and runtime

• 201 focused software tests passed, including 23 new closure/admission cases and the previous registration, correction, route search, causal recovery, reference version and session-default suites. One existing Starlette/httpx deprecation warning remains. Ruff and diff whitespace checks passed.
• Final real-source qualification uses the same immutable small 579.797 m and large 1595.409 m recordings and their sealed decode caches. No new scanner recording, hardware command, frontend change or service restart is required.
• The shared policy retains the small 20/5-second closure and selects 40/30 seconds for the large ring. Source frame holdout remains correlated by the upstream K1 map and is not independent survey truth.
• Final same-source local checks pass 52/52 on the small ring and 136/136 on the large ring. Frozen seam holdout overlap is 89.434% and 95.093%, respectively.
• Repeating the independent full small-ring planner capture against the final rebuilt small atlas restores 101/101 numerical and causal acceptances, matching the 101/101 frozen-reference baseline. This is a warm map-version regression, initialized by its existing old-map cold fit, not another cold-start claim.
• Complete large-map atlas and trajectory arrays equal the previously tested diagnostic atlas. Its independent short-pass 29/29 and cold numerical search evidence remain applicable by exact identity, not by assuming similar maps have identical behavior. No full independent pass through the new territory is inferred from that reuse.
• A full numerical search for the planner's ja-sun-009-100m was stopped after host swap increased. Its exact temporary worker was terminated and reaped; canonical service 8000 was retained. This incomplete run is neither a localization success nor a geometric rejection. No repeated heavyweight search is required to establish equivalence of identical map arrays.

Private evidence is under .runtime/audits/2026-09-21-closure-v2/, with the initial candidate retained at the root and final cascade results separately under cascade/. Earlier observations and source recordings remain immutable.

Reviewed version admission

The existing packager rechecked all 13,608 full-resolution large-map frames and all 13,624 poses against the frozen field, plus source transport/clocks and the complete review seal. It published a separate content-addressed full map. The ordinary large-session selection now resolves to its corrected generation and 1599.807 m in the canonical planner API. The small-session default pointer is unchanged; existing planner studies retain their pinned generations. Raw captures remain available, and no vehicle-control authority is granted.

Canonical API acceptance confirms that the spatial overview defaults to corrected and uses the same admitted map generation as ordinary planner selection. The original-generation-pinned view remains original. Both Original and Corrected preview RRDs were rendered successfully at the 80 m height ceiling. Browser visual QA and a new full playback export were not run. The final private evidence seal rechecked 212 input files with no changes, no producer changes, and exact small/large geometry equivalence confirmed.

The operator workflow remains explicit: reconstruct → frozen review → package → publish paired overview → activate the session default. No new UI or background automatic capture correction has been introduced. Publication and admission reuse the existing ordinary-name selection and Original/Corrected inspector.

Decision and remaining scope

The shared cascade is qualified for these two known start-area revisits, with explicit fail-closed acquisition/review boundaries. Universal means one policy and regression corpus, not a claim that every possible ring will close.

Independent new-territory repeat accuracy, absolute geometry, the actual spatial distribution of drift, arbitrary crossing-loop graphs, seasonal transfer and onboard/live timing remain unproven. One accepted seam plus a smoothness prior cannot establish those facts. Incomplete optional planner searches remain open; do not relabel them as passed or use mixed-coverage loops as an aggregate score.