Node extraction quality: expose hypothesis-strength ratio + deterministic tie-break

[rolker]

Summary

From the #30 comparison against Calder's original CUBE. Two node-extraction
quality improvements: surface the hypothesis-strength ratio Calder computes, and
make hypothesis tie-breaking deterministic.

Part of #30.

1. Expose the hypothesis-strength ratio (real diagnostic value)

Calder's depth extraction returns a hypothesis-strength ratio — a measure of
how dominant the chosen hypothesis is over the runners-up
(max(0, MAX_HYPOTHESIS_RATIO - N_best/N_rest), cube_node.c:1566-1586,
:2108-2112). It does not change the chosen depth; it is a node-ambiguity
signal. The port's extractDepthAndUncertainty (node.cpp:167-187) drops it.

Exposing it (e.g. as a third grid layer alongside elevation/uncertainty) would
let CAMP/rviz flag cells where multiple competing depth hypotheses survive —
useful for QA and for steering re-survey.

2. Deterministic tie-break

Calder stores hypotheses in a push-down linked list (newest at head); the port
uses a std::vector with push_back (newest at tail), reversing iteration. On an
exact error / num_samples tie the two pick different hypotheses
(bestHypothesis node.cpp:65-79, chooseHypothesis :189-200). Calder's own
qsort comparator never returns 0, so its tie outcome is also undefined.

Exact float ties are vanishingly rare in practice, but a defined rule makes the
behavior testable. Add an explicit tie-break (e.g. on a tie, prefer higher
num_samples, then lower current_variance, then most-recent) and a test that
pins it.

Acceptance

• Hypothesis-strength ratio computed and surfaced from node extraction (+ optional
  grid layer); test pins the ratio for a known multi-hypothesis node.
• Deterministic tie-break rule implemented in both selection paths; test asserts
  the chosen hypothesis on a constructed exact tie.

---

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