Context

Retroactive audit of merged PR #1191 (two-tier pass execution, commit e1b3617) surfaced a DRY miss.

DRY: Four copies of child-slot iteration

The tuple `("input", "left", "right", "subquery")` enumerates the structural child slots of every `LogicalPlan` subclass. It currently appears in four places:

1. `graphistry/compute/gfql/passes/unnest_apply.py:~40` — `_unnest_tree` iteration
2. `graphistry/compute/gfql/passes/predicate_pushdown.py:~45` — `_rewrite_tree` iteration
3. `graphistry/compute/gfql/ir/verifier.py:30` — `_CHILD_SLOTS` constant
4. `graphistry/compute/gfql/ir/physical_planner.py:190-194` — `PhysicalPlanner._children`

Each uses the same pattern: try each slot on the current plan node, rebuild with `dataclasses.replace(..., **children_updates)`, preserve child identity when unchanged.

Adding a new `LogicalPlan` subclass with a novel child slot (e.g., a second `input` or an alternative arm) requires editing all four sites. The test suite might not catch a miss if the new slot's traversal is exercised in a pass that skips the missed site.

Proposed fix

Introduce a shared helper in `graphistry/compute/gfql/ir/logical_plan.py`:

```python
CHILD_SLOTS: Tuple[str, ...] = ("input", "left", "right", "subquery")

def rewrite_children(plan: LogicalPlan, rewrite: Callable[[LogicalPlan], LogicalPlan]) -> LogicalPlan:
"""Apply rewrite to each child slot and return a new plan, preserving identity when unchanged."""
updates: Dict[str, LogicalPlan] = {}
for slot in CHILD_SLOTS:
child = getattr(plan, slot, None)
if isinstance(child, LogicalPlan):
rewritten = rewrite(child)
if rewritten is not child:
updates[slot] = rewritten
return replace(plan, **updates) if updates else plan
```

Replace the four duplicated sites with calls to this helper. The `CHILD_SLOTS` tuple becomes the single source of truth.

Related gap

No regression test locks the `rewritten is not child` identity-preservation invariant for deep trees (mentioned in `test_pass_manager.py`'s convergence logic — if identity breaks, `changed` signaling breaks). Add one after consolidation.

Priority

p3 — cleanup; does not block any feature. Worth doing before the M4 pass framework grows more passes.