Function select_native_points 

pub fn select_native_points(
    b: &[u8],
    width: usize,
    height: usize,
    region: Rect,
    mask: Option<(&[u8], u8)>,
    guide: Option<(Gray<'_>, usize)>,
) -> Vec<i64>

Picks the native-refinement sample points for b inside region: a step×step cell grid, keeping one point per cell rather than sampling only the pixel at the cell’s corner (its lattice node). A node-only sampler finds an edge with probability proportional to local edge density, so a sparse page (few edges anywhere) starves the sample (< 12 points → Infinity, the true step read as “static”) and a screen-fixed line that happens to align with a lattice row or column dominates it (every node on that row is an edge, nothing else is) — see docs/ARCHITECTURE.md §五/§十. Depends only on b, region, mask and guide, not on any hypothesis, so a caller evaluating several candidate displacements can select once and reuse the points for all of them.

Without a guide: a full-resolution scan keeps, per cell, the pixel with the strongest native luma gradient (ties broken by row-major order) — the exact per-cell argmax.

With guide = Some((g, f)) — g is b’s own analysis-scale gray at integer downscale factor f (raster::downscale_gray; analysis pixel (ax, ay) covers the native block [ax·f, ax·f+f) × [ay·f, ay·f+f), edge blocks partial): coarse-to-fine instead of full-resolution. Per cell, first pick the analysis pixel (among those whose native block intersects the cell, skipping the analysis image’s own 1px border and any pixel whose block centre fails mask) with the strongest analysis-scale gradient, then the native pixel with the strongest native gradient inside that one analysis pixel’s block ∩ cell ∩ window ∩ mask — luma computed only for that handful of pixels and their four neighbours, not a full luma plane. f = 1 makes g b’s own native luma one-for-one (downscale_gray’s factor == 1 branch), every block a single pixel, and this reduce exactly to the unguided per-cell argmax — this is a property relied on elsewhere (the f = 1 odometry parity cases), not merely incidental.