fix: zero-init TGS empty_trigrid_node_ + AOS Point/AOSNode (root-cause CI flakes)

[LimHyungTae]

Two related uninit-memory bugs that caused 1-1.5% run-to-run mismatches in the Python smoke tests on the GitHub Ubuntu runners. Same shape in both: a class member declared without an initializer, copied into a working buffer that the algorithm then assumes is zeroed.

TGS

initTriGridField had four no-op expression statements where it clearly intended to zero-initialize the Eigen members of empty_trigrid_node_:

empty_trigrid_node_.normal;            // evaluates, drops the lvalue
empty_trigrid_node_.mean_pt;           // same
empty_trigrid_node_.singular_values;   // same
empty_trigrid_node_.eigen_vectors;     // same

Eigen vectors/matrices don't zero-init by default. empty_trigrid_node_ then got copied into every cell of trigrid_field_, seeding the whole field with stack garbage.

Fix: replace with explicit Eigen::Vector3f::Zero() / Eigen::Matrix3f::Zero().

AOS

class Point declared int label; (and float x,y,z;, uint idx;) without in-class initializers. range_mat_ entries got assigned fresh Point copies in sphericalProjection, leaking the uninit label field into the working set. class AOSNode had the same problem on its start / end members. Combined, they made cluster filtering decisions in labelPointcloud run-to-run nondeterministic — about 1-1.5% of points flipped between "in some cluster" and "no cluster" on consecutive segment_objects calls with identical input.

Fix: brace-init (label{0}, start = 0, etc.).

Why this is safe under the "never break the pipeline" rule

This is a correctness fix, not a behavior change. The existing code's output was already nondeterministic, so there's no single "original behavior" to preserve. What we want to preserve is the deterministic majority result, which is exactly what the algorithm produces in the common (non-uninit-read) path.

Verified on macOS arm64:

$ ./build/tests/regression_kitti kitti00_000000.bin /tmp/run.bin cpp/tests/data/kitti00_000000_gold.bin
REGRESSION PASS: bit-identical to cpp/tests/data/kitti00_000000_gold.bin (374008 bytes)

So the algorithm's output on KITTI sequence 00 frame 0 (with AOS seed=42) is byte-for-byte the same as the gold dump captured pre-fix. The fix moves the previously-stochastic minority into the same canonical answer the deterministic majority was already producing.

test_xyz_and_xyzi_inputs_both_accepted and test_class_api_matches_convenience_function ran 30 times back-to-back on macOS — 30/30 passed for both (each was flaking ~1-1.5% on the Ubuntu runner before this commit).

Full pytest python/tests/ -> 7/7 passed.

Test plan

• CI green across all 6 jobs (KITTI bit-compare in particular).
• Two CI runs back-to-back (workflow re-run from this PR's HEAD) both green — confirms the AOS-side flake is gone, not just rerun-lucky.