GNC/guidance/target_finder.py — FOV coverage & ANCHOR scoring

EETargetFinder: marks per-triangle FOV coverage and scores candidate camera poses. Coverage is trajectory-driven — the scorer is inert in POSE (the adopted default).

Per-triangle hit count

At each pose, every triangle struck within the working depth band is counted:

$$h_{\text{tri}} \leftarrow h_{\text{tri}} + 1$$

$h$ is the shared ground truth for both coverage and novelty — one counter, two uses.

Area coverage — eq 7.1

Let $\mathcal{S}$ be the set of triangles seen at least once. Area coverage is:

$$\text{coverage} = \frac{\displaystyle\sum_{i \in \mathcal{S}} A_i}{\displaystyle\sum_j A_j}$$

$A_i$ = triangle area. Independent of candidate sampling — driven by orbit traversal.

Novelty score

Per-triangle novelty from the hit count $h$:

$$\sigma_n = \frac{1}{1 + h}$$

Fresh triangle ($h=0$): $\sigma_n = 1$. Repeatedly seen ($h \to \infty$): $\sigma_n \to 0$.

ANCHOR weighted-product score — eq 7.2

Candidate (target, pose) pairs are ranked by:

$$\Omega = \prod_i \sigma_i^{w_i}$$

$\sigma_i \in [0,1]$: view quality, novelty $\sigma_n$, area, stability, motion, anchor, manipulability. Weights $w_i$ active only in EXPONENTIAL mode — inert in POSE.

View-quality sub-scores

Three geometry terms folded into $\sigma_{\text{view}}$:

$$q_1 = \text{dist ramp}(d, D_{\min}, D_{\max}) \qquad q_2 = \cos\theta_{\text{inc}} \qquad q_3 = 1 - \frac{d_{\text{pix}}}{d_{\text{pix,max}}}$$

$$\sigma_{\text{view}} = \frac{w_d q_1 + w_\theta q_2 + w_\xi q_3}{W_{\text{Kong}}}$$

Coverage: orbit traversal, not selection

In POSE mode the scorer never runs — yet:

$$\text{coverage} \xrightarrow{\text{full helix}} 1.0$$

The lever is the orbit path (pole-to-pole helix), not scoring weights or reselect cadence. Achieved: coverage $= 0.9990$ at the adopted operating point.