fix(query): correct frontier eviction in shortest path priority queue

[shaunpatterson]

Summary

• Replace pq.Pop() with removeMax() — the old code removed the last element of the underlying slice (arbitrary in a min-heap) instead of the highest-cost node. removeMax() does a linear scan and uses heap.Remove() to properly evict the least promising candidate while maintaining the heap invariant.
• Push before evict — the old code evicted before pushing the new node, so the newcomer was never considered for eviction. A high-cost new node would be admitted while a lower-cost existing node was dropped. Now we push first, then evict the max from the full candidate set.
• Guard MaxFrontierSize > 0 — skip eviction when the limit is disabled (zero value), preventing accidental eviction when no limit is configured.

Builds on the fix in #9607 which identified the pq.Pop() vs removeMax() issue, and addresses two additional correctness problems (evict-before-push ordering and boundary condition) found during review.

Test plan

• 24 unit tests with 100% code coverage on all priority queue functions (indexOf, Len, Less, Swap, Push, Pop, removeMax)
• 10 integration tests exercising maxfrontiersize in DQL shortest path queries:
  • Optimal path found with tight frontier
  • maxfrontiersize=1 doesn't crash (single-path and k-shortest)
  • Large frontier produces identical results to unconstrained query
  • Linear chain graph with small frontier
  • Key regression: high-cost node evicted instead of low-cost node on optimal path
  • Key regression: push-then-evict ensures newcomer competes for eviction

🤖 Generated with Claude Code