Two latent traps in the first-pass centroid scan:
- A bbox where x/y were finite but z was NaN would skip the z update;
if NO entity contributed a finite z, `min[2]` stayed at `INFINITY` and
`max[2]` at `NEG_INFINITY`. `(INF + NEG_INF) * 0.5 = NaN`, so the defn's
`local_offset.z` was NaN, propagating to every cast point in the defn.
Track `have` per axis and default missing axes to 0.
- A single corrupt entity at parser-garbage coords (1e15+) silently
dragged the centroid out of the real geometry's neighbourhood. Reject
per-axis values above `SANE_EXTENT` so one bad bbox doesn't ruin the
whole defn.
Also include `Insert::insert_point` as a contributing sample so a
composite block (only nested Inserts, no direct wires) still gets a
non-zero offset — useful for the nested-defn AABB transforms even
though no direct LocalWire stores against it.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
0d543e8c88