Lists: one allocation per list again; [each acc, x] appends in place

[revarbat]

Refs BelfrySCAD/BelfrySCAD#668

This follows up 1.35.0's shared-prefix lists (#225). That change made accumulator loops linear but, measured properly, cost ordinary BOSL2 code 1–5.6% in instructions and inflated a long docs build's footprint. This PR keeps the linear appends and removes both costs. It also extends in-place appending to [each acc, x].

Why 1.35.0 regressed (profiled with sample; heap inspected with heap/vmmap):

• One extra allocation per list. A ValueList held a shared_ptr to a separately allocated ListBuffer, which held the std::vector. That showed up as about +2.5% in malloc/free and +0.7% tearing down the ListBuffer control block.
• ×2 growth with at least 16 slots produced large freed blocks. macOS keeps those resident: after the shapes3d docs the live heap was unchanged (74.6 vs 74.4 MB), but 305 MB of freed large blocks were still held.

Changes:

• makeList: the list and its buffer header share one block, and the element vector is adopted, not copied. That's the same two allocations a plain vector cost.
• Cached element pointer: each list caches it, so a read is one hop.
• Appended lists: a list appended in place is a small ValueList viewing its parent's buffer, keeping the parent's block alive through a shared_ptr that points into it.
• Construction: ListItems is neither copyable nor movable, so all list construction goes through makeList, which the compiler enforces.
• Growth is ×1.5 with no minimum.
• [each acc, x]: a list literal or comprehension whose first contribution is each <list> now starts as that list (ListBuilder), on the VM and the interpreter. It was still quadratic: 50k steps took 6.1 s; 400k now take 0.2 s.

Measured with instructions retired (/usr/bin/time -l), which load doesn't affect. Wall-clock timings on this machine were distorted by unrelated load, and that's how #225 shipped a false claim.

Nine BOSL2 scripts, against 1.34.0:

script	1.35.0	this PR
paths	+5.6%	+0.8%
vnf	+3.5%	+0.2%
skin	+2.6%	+0.5%
rounding	+2.5%	+0.1%
cuboids	+1.8%	+0.5%
threads	+2.2%	−0.9%
sweep	+1.3%	−0.4%
text	+1.3%	+0.9%
gears	+1.0%	+0.1%

Bytes allocated: 1.35.0 was +1.4% to +33.5% over 1.34.0; this PR is +1.6% to +24%.

Full BOSL2 docs build:

	instructions	cycles	peak footprint
1.34.0	7,522.6 G	1,862.2 G	1,831 MB
1.35.0	7,753.8 G (+3.1%)	1,915.9 G	1,784 MB
this PR	7,526.2 G (+0.05%)	1,861.9 G	1,651 MB

shapes3d docs, after the run:

• 1.34.0: footprint 380 MB, 3.6 MB of freed large blocks held.
• 1.35.0: 738 MB footprint, 305 MB held.
• This PR: 390 MB footprint, 3.6 MB held.

Append loops are still linear: a million concat appends take 0.6 s, and #668's script stops with "Recursion detected" in under 2 s.

Tested:

• ctest: 2014 passed.
• AddressSanitizer: all 1355 tests, plus the append, #668 and BOSL2 scripts, clean.
• macOS leaks: 0.
• test_python_bindings.py: 61 passed, on a wheel from this branch.
• BelfrySCAD suite: 2193 passed.
• BOSL2 docs A/B against 1.35.0: 56/56 markdown files identical and logs identical. 2636/2640 images are pixel-identical. The other 4 are cubetruss images with 1–28 edge pixels changed. Those come from Patch Manifold v3.5.2: fix a heap overflow in CsgLeafNode::Compose #227's Manifold patch, not from this PR: an unpatched build of this branch gives cubetruss meshes identical to 1.35.0's.

🤖 Generated with Claude Code