Motivation

• Reduce repeated ngv.query_nearest(...) lookups and avoid rebuilding sample vectors for each radius step to cut CPU overhead when fitting normals.
• Trade a bounded amount of per-sample memory for large preprocessing savings when adaptive radius expansion revisits overlapping neighborhoods.

Description

• Add a per-thread cache keyed by a packed (plane_idx, slice_idx) for ngv.query_nearest(...) results and a query_nearest_cached(...) helper to reuse resolved grids across radius steps and nearby samples.
• Introduce SegmentCandidate to store per-segment metadata (dir_unit, delta_xyz, weight, dist2, is_short_path) and add gather_segment_candidates(...) that collects and sorts all candidates once for the maximum radius.
• Refactor radius handling so gather_samples_for_radius(...) advances per-plane cursors through the distance-sorted candidate lists and appends only newly in-range segments for larger radii, rather than rebuilding vectors from scratch.
• Extend per-thread FitBuffers to persist candidates and per-plane candidate_cursor so candidate lists and cursors survive across radius iterations inside a sample evaluation.
• Preserve existing adaptive-radius logic and fitting call sites (the solver uses the same dirs_unit, weights, and deltas_xyz inputs after the refactor).

Testing

• Attempted to configure the project with cmake -S volume-cartographer -B volume-cartographer/build -DCMAKE_BUILD_TYPE=Release, but the build configuration failed because dependency fetches from GitHub were blocked (CONNECT tunnel failed, response 403 for xtl.git), so a full compile/run verification could not be completed.
• No automated unit tests were executed in this environment due to the configuration/build failure; changes were validated by static inspection of the modified source to ensure API usages match the original flow.

Codex Task