Fix non-deterministic caller resolution in call graph tracking

[Copilot]

On second rerun, get_rays() fails with assertion error despite identical inputs. Root cause: caller params_hash resolution iterates all graph nodes without breaking, using last match. When multiple nodes exist for the same function (after 2nd+ reruns), selection becomes non-deterministic.

Changes

• Added execution stack to track active call chain in CallTracker

  • Push callee node on function entry
  • Pop on exit via try-finally for proper cleanup

• Modified caller resolution to check stack first (top = current caller), fallback to graph search with break

• Fixed caching logic to allow None return values (removed incorrect result is not None check)

Before/After

# Before: Non-deterministic, uses last node in iteration order
caller_params_hash = -1
for n in self._call_graph.nodes:
    if n.func == k:
        caller_params_hash = n.params_hash  # No break!

# After: Deterministic, uses active caller from stack
if self._execution_stack:
    caller_node = self._execution_stack[-1]
    if caller_node.func == k and caller_node.module == caller_module:
        caller_params_hash = caller_node.params_hash

The stack ensures correct caller identification when multiple nodes with same function exist in the graph.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.