Port Calder's predicted-surface / guided-disambiguation subsystem (prior-depth offset + CUBE_PRIOR/POSTERIOR)

[rolker]

Summary

Tracking issue (from the #30 comparison against Calder's original CUBE). Calder's
predicted-surface / spatial-guide subsystem was not ported. The current port
runs pure data-driven CUBE. This is fine today but is a known gap to port when
a prior/predicted surface becomes available.

Part of #30. Feeds / fed by the chart-prior importer #44 and the bathymetric data
store epic rolker/unh_marine_autonomy#86.

What's missing

The port has a predicted_depth_ member (node.h:198) initialized to
INVALID_DATA and no setter — it is read in three guards but never assigned a
real value. Consequently:

1. Prior-depth propagation offset (node.cpp:116-124, commented out) is moot —
   its precondition predicted_depth_ != no_data_value can never be true. In Calder
   the node's pred_depth is set from a predicted/prior surface
   (cube_node.c:1091) and used to refer a sounding to that surface
   (offset = pred_depth - range) and to bound blunders / capture.
2. Guided disambiguation — Calder's CUBE_PRIOR (closest-to-guide) and
   CUBE_POSTERIOR (guide-weighted log-posterior) hypothesis selectors
   (cube_node.c:2037, :2136) are not ported; only the data-driven most-samples
   selector + nomination override exist. The unused DEFAULT_EXTRACTOR = CUBE_LHOOD
   constant reflects this gap.

These are one subsystem: both need an external predicted/prior surface.

Scope when picked up

• A mechanism to supply a per-node predicted depth + variance (from a chart prior
  / GGGS bathy store / previous-epoch surface).
• Wire the propagation offset and predicted-surface blunder/capture bounds.
• Port CUBE_PRIOR / CUBE_POSTERIOR extractors and make the disambiguation
  method selectable.
• Tests covering guided init, propagation, and prior/posterior selection (the
  comparison agents drafted a canonical guided-vs-most-samples regression case).

Not in scope now

No code changes in the near term — the current pure data-driven behavior is
documented as an intentional divergence (#30 doc). This issue exists so the gap is
on the roadmap rather than buried.

---

Authored-By: Claude Code Agent
Model: Claude Opus 4.8 (1M context)