Reference blunder gate has no coverage where the prior is NoData — extreme false-deep outliers survive at lake edges

[rolker]

Problem

The 2026-07-02 authoritative Massabesic rebuild (unh_echoboats#353: 36 bags, 11.0 M pings, reference prior active via --reference-store, #89/#96 blunder gate) still contains 2,358 impossible-depth survey cells (0.027%) — up to 4,354 m "deep" in a 17 m lake. The population is nearly identical to the pre-gate legacy store (2,654 cells / 0.037%), i.e. the reference blunder gate barely touched the extreme false-deep detections it was built to reject.

Likely mechanism (to verify)

The worst tiles (e.g. 10_17791_13902) sit at the lake edge/dock areas. The reference layer is rasterized from the NH GRANIT depth-band polygon union — shoreline nooks, dock aprons, and anything outside the polygon union have NoData reference cells, and a NoData prior means no gate for soundings landing there. So exactly where multipath/false detections are densest (dock structures, shallow edges), the gate is absent.

Ask

1. Confirm the surviving outliers correlate with NoData reference coverage (quick join of outlier cell coords vs reference validity).
2. Decide the fallback policy for gate-less cells — e.g. a store-wide max-plausible-depth bound from the reference layer's own statistics (deepest reference band + margin), or gate against the nearest valid reference cell within N cells.
3. Regardless of gating: these are upstream false detections (M3 detector artifacts); the durable fix may belong in detector-side filtering (flags/quality thresholds in the projector) rather than only in CUBE's prior gate.

Context

• Same outlier set was present in the June incremental store (noted during cube#87 work: "false detections placed too-deep — handle upstream separately").
• Not a regression of Adopt the simplified cube stores: tile-open seed precedence, batch regen, sufficient-statistic backscatter #96/import_bag: seed CUBE predicted surface from the Chart (contour) prior to reject false-deep detections #89 — the rebuild slightly reduced the population — but the gate's coverage hole is now measured and worth closing while the survey data is fresh.

---

Authored-By: Claude Code Agent
Model: Claude Fable 5