Start a named draw group from an empty hold, and state the decision's goal and cost exactly

docs/decisions.md

@@ -244,13 +244,12 @@ carry both rules and both scopes anyway, and tell them apart at every step.
 
 ## A hold set makes its draw maps on the first read
 
-Every render starts from the zero `holdSet`, per goal 2's one spelling. Making the maps
+Every render and every draw group starts from an empty hold, since goal 2 prices a reader
-where the set is declared kept a record's on its frame's stack, but a `Fake`'s set lives
+meeting two spellings of one empty set. Making the maps where the set is declared kept a
-on the `Generator`, so there it cost every render two heap allocations and about half the
+record's on its frame's stack, but a `Fake`'s set lives on the `Generator`, so there it
-cheapest one's time, measured 2026-09-28. Lazily, a render reading a reference path pays
+cost the cheapest render two heap allocations and 240→390 ns, measured 2026-09-28.
-those two allocations, a record or struct included, and one reading none pays nothing;
+Lazily, a render reading a reference path pays those two allocations, a record or struct
-goal 11 holds either way. The allocation gate over a repeat of a reference path and over
+included, and one reading none pays nothing; goal 11 holds either way.
-a named draw group prices that, and pins what keeps a hold set off the heap.
 
 ## A category never references itself, and a record's fences run at load
 

holdset.go

@@ -63,7 +63,7 @@ func (sc renderScope) hold() *hold {
 		if sc.set.named == nil {
 			sc.set.named = map[string]*hold{}
 		}
-		d = &hold{variant: map[string]node{}, value: map[string]readValue{}}
+		d = &hold{}
 		sc.set.named[strings.Clone(sc.group)] = d
 	}
 	return d

todo.md

@@ -300,9 +300,6 @@ because pairing a street with its exact postnummer rewrites shipped rows.
   flag, an exit code, an exported name, a fence, a builtin and the lowest Go.
 - Measure the allocation gate over the shipped `geo` trees too: `pinSet.inline [8]` is sized
   for them, and `perf_test.go` checks a synthetic five-deep tree only.
-- Make a named draw group's hold lazy, `&hold{}` in `renderScope.hold`: it escapes
-  to the heap anyway, so its two eager maps buy no stack and cost two allocations per
-  group that may never read through them. Check it against the alloc tests.
 - Name goal 2's four criteria with the axes the panel scores, so a recorded score maps
   back to the clause it came from: the goal spells them out in prose while `todo.md`
   records `Navigation`, `Locality`, `Shape` and `Self-sufficiency`.