Extend Op::PushBuiltinWrap to hull/minkowski/render/extrude/offset/roof

[revarbat]

Summary

• Every remaining non-leaf builtin whose own children can wrap a recursive call still fell to Op::NativeStatement — one real native C++ reentry per level, capped at kMaxDriveVmNativeDepth=40 (same risk Op::PushBuiltinWrap Close the NativeStatement gap for translate/color/modifier-wrapped recursion #63 and Op::PushCsgWrap Add Op::PushCsgWrap: eliminate union/difference/intersection's native reentry #67 already closed for their own subsets).
• hull()/minkowski()/render()/linear_extrude()/rotate_extrude()/projection()/offset()/roof() all share resolveTransform's own "resolve args, evaluate children, build params" shape, so they reuse Op::PushBuiltinWrap's existing bracket via 6 new Kind values instead of a new opcode set — no new mechanism, just widening the one that already exists.
• One behavioral subtlety: roof()'s "Unknown roof method" warning is an observable side effect, so its params are computed at Pop time (after children) instead of Push time like every other kind here, to preserve native's own echo/warn ordering exactly. Verified with a dedicated ordering test.
• intersection_for remains uncovered — its group_sizes are keyed by runtime loop iteration count, not a static source-statement list, so it needs real loop-compilation machinery neither existing bracket shape provides. Flagged as a separate, larger follow-up (not silently skipped).

Test plan

• Full 737-test suite green both OSCAD_BYTECODE_VM on (default) and off, up from 727 (10 new tests: one recursion-depth test per builtin, $-arg propagation into a compiled wrap, and the roof warn-ordering test).
• Caught and fixed a test-design bug during verification: offset(r=1) nested 1500 levels deep is a real, unbounded geometric operation (each level's rounded-join offset grows both size and vertex count) — hung for minutes before being rewritten to offset(delta=0) (a real no-op at every level, same opcode path, none of the compounding cost).
• Real-world smoke check: BOSL2's examples/fractal_tree.scad renders to a byte-for-byte identical STL before/after this change.
• Version bumped to 0.13.6.

🤖 Generated with Claude Code

[revarbat]

Verification follow-up: while working on the intersection_for follow-up (#69), found that my local CLI binary (tools/cli/openscad-cpp-evaluator) hadn't been relinked since the start of the session — a separate build target from the test binary, easy to miss. This meant the fractal_tree.scad byte-identical check in this PR's description ran against a stale binary. That check was weak anyway (that script doesn't exercise hull/minkowski/render/extrude/offset/roof — it's mostly cylinder/attach/rotate/color), so it wasn't invalidating anything, but it also wasn't testing what it claimed to.

The unit test suite itself was never affected (CMake always relinks oscad_eval_tests correctly on --target oscad_eval_tests), so the 737/737 differential result stands. Re-ran the actual depth check with a properly rebuilt CLI to be sure:

hull depth=39: SUCCEEDED   (was capped at 39 before this fix, per the earlier investigation)
hull depth=1500: SUCCEEDED

Confirms the fix is real, not just unit-test-shaped. No code changes needed here — just flagging the gap in case CI's own from-scratch build didn't already make this moot (it should have).