FloydWarshallOptimized: `next` matrix can contain a cycle for a reachable pair (root cause behind the v0.7.1 hang guard)

[apotema]

Background

v0.7.1 (#12) added a defensive cap to setPathWithMapping{,Unmanaged}: the next-hop reconstruction loop was infinite-looping because the next matrix contained a cycle that never reaches the goal, growing the path list unbounded → hard hang + OOM. Surfaced loading a 1000-worker flying-platform colony; a stack sample pinned it to getPath looping with ensureTotalCapacity climbing every iteration.

The cap makes reconstruction safe (returns NoPathFound past size hops), but it papers over the real defect:

The real question

A correct Floyd-Warshall next matrix must never contain a cycle for a pair where hasPath == true. next[u][v] is the first hop on a shortest path u→v; following it must strictly make progress toward v and terminate in ≤ size hops. A cycle means dist and next are mutually inconsistent — i.e. generate() produced a corrupt matrix, OR the input graph has a property the optimized variant mishandles.

Hypotheses to investigate

1. Optimized (SIMD/parallel) generate() bug — the next update during relaxation is wrong/raced for some node ordering, producing a next that doesn't match the final dist. Compare FloydWarshallOptimized vs the scalar FloydWarshall on the same graph and diff the next matrices.
2. Equal-cost / zero-weight edges — ties in relaxation (dist[i][k]+dist[k][j] == dist[i][j]) updating next inconsistently, creating a 2-cycle (next[u][v]=w, next[w][v]=u at equal cost).
3. INF / disconnected handling — a pair marked reachable (hasPath) whose next chain routes through a wrongly-relaxed entry.
4. Mapping layer (nextWithMapping/reverse_ids) — a stale/duplicate id mapping after the graph is rebuilt (the flying-platform graph is rebuilt on scene load), so the index→id translation yields a cycle even if the raw matrix is fine.

Repro lead

Deterministic: load big_colony (1000 workers) in flying-platform on the bgfx backend; the post-load repath storm hits the cyclic pair. Better: capture the exact (start, goal) node ids + the graph (edges) at the moment the guard fires, then reproduce in a zig-utils unit test by building that graph and asserting generate() yields an acyclic next.

Acceptance

• A minimal failing case (graph + start/goal) reproduced in a zig-utils test.
• Root cause identified (one of the above or new).
• generate() fixed so next is always consistent with dist; the v0.7.1 cap stays as defense-in-depth.

Related: #12 (the guard), flying-platform#540 (pin bump).