openscad-cpp-evaluator 0.13.3__tar.gz → 0.13.5__tar.gz

{openscad_cpp_evaluator-0.13.3 → openscad_cpp_evaluator-0.13.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: openscad_cpp_evaluator
-Version: 0.13.3
+Version: 0.13.5
 Summary: C++ OpenSCAD evaluator with Python bindings
 Author: NeuSCAD Contributors
 License-Expression: MIT

{openscad_cpp_evaluator-0.13.3 → openscad_cpp_evaluator-0.13.5}/include/openscad_cpp_evaluator/bytecode.hpp

@@ -325,10 +325,148 @@ enum class Op {
     // same seam, now a counter).
     PopBuiltinWrap,
 
-    // A single "native passthrough" statement -- intersection_for,
-    // children(), union/difference/intersection and every other builtin
-    // module call NOT covered by Op::PushBuiltinWrap (see that op's own
-    // doc comment for exactly which builtins ARE covered), the `*`
+    // -- CSG-wrap compilation (closes the LAST native-reentry source in ----
+    // -- the original NativeStatement gap: union()/difference()/           -
+    // -- intersection()) ----------------------------------------------------
+    // union()/difference()/intersection() weren't covered by the original
+    // Op::PushBuiltinWrap (see that op's own doc comment) because they need
+    // bespoke bookkeeping PushBuiltinWrap's single all-children bracket
+    // doesn't do: Evaluator's own resolveCsg (booleans.cpp) evaluates each
+    // TOP-LEVEL child statement of the block SEPARATELY and records how
+    // many CSGNodes it individually contributed ("group_sizes") -- e.g.
+    // `difference(){ A; B; C; }` = A - (B u C), preserving A's own group
+    // even when A itself expands to more than one body (an attachable()-
+    // style call returning parent+children as one operand). A flat
+    // all-children bracket the way Transform/Color/Modifier use would lose
+    // that grouping entirely. This op pair (PushCsgWrap/PopCsgWrap) plus
+    // Op::CsgGroupStart/CsgGroupEnd (below) replicate resolveCsg's exact
+    // two-pass shape (all assignments first, THEN one evalChildren-per-
+    // geometry-statement) as inline compiled bytecode instead, exactly as
+    // PushBuiltinWrap already did for translate/rotate/scale/mirror/
+    // multmatrix/resize/color/#/%/! -- same rationale (a recursive
+    // union()/difference()/intersection()-wrapped module chain was still a
+    // real Windows native-reentry depth risk, just never independently
+    // fixed when PushBuiltinWrap's own set was).
+    //
+    // a = index into CompiledChunk::csgWrapSites. Runtime handler: captures
+    // ev.randsCallCount() BEFORE argument resolution (same rands-in-args
+    // taint reasoning as PushBuiltinWrap), resolves the (rare, $-only)
+    // arguments via resolveCallArgs exactly like resolveCsg itself does
+    // (discarding the positional/named result -- union/difference/
+    // intersection take no real parameters -- keeping only the possibly-
+    // $-scoped child ctx), pushes that ctx onto f.ctxChain unconditionally
+    // (mirrors Transform/Color's own unconditional push -- a bare
+    // `union() {...}` with no `$fn=...` override still pushes a ctx that's
+    // merely a copy, cheap and uniform rather than a special-cased branch),
+    // pushes a fresh ev.treeStack_ frame (every child statement's own
+    // CSGNode(s), across every group, land flat in this ONE frame --
+    // mirrors Evaluator::buildTreeNode/evalModularCall's own single
+    // treeStack_.emplace_back() around the whole call, not one per group),
+    // and stashes {op, randsBefore, siteIdx, empty groupSizes} onto
+    // VmFrame::csgWrapStack (a real per-frame LIFO, same reasoning as
+    // builtinWrapStack: nested/sequenced CSG wraps within one frame's own
+    // instruction stream, e.g. `union() { difference() {...} }`). The
+    // compiler always emits a plain Op::CheckDebugStatement immediately
+    // before this (see emitCsgWrap, bytecode_compiler.cpp), mirroring
+    // emitBuiltinWrap's own pattern -- this is a genuine statement doing
+    // real work here, not a call transferring control to a declaration.
+    PushCsgWrap,
+
+    // Opens one "group" within an already-open Op::PushCsgWrap bracket --
+    // emitted immediately before each top-level GEOMETRY child statement's
+    // own inline-compiled bytecode (compileStatementList of exactly that
+    // one statement). Records ev.treeStack_.back().size() into
+    // csgWrapStack.back().groupStartSize -- always operates on the
+    // TOPMOST (innermost still-open) csgWrapStack entry, matching
+    // PopBuiltinWrap's own back()-is-always-mine LIFO discipline, so no
+    // operand is needed. Assignments among the block's children are
+    // compiled separately, BEFORE any CsgGroupStart/End pair at all (see
+    // emitCsgWrap) -- mirrors resolveCsg's own two-pass split exactly
+    // (Evaluator::evalChildren(assignNodes, effCtx) always runs to
+    // completion before the per-geoNode loop starts).
+    CsgGroupStart,
+
+    // Closes the matching Op::CsgGroupStart: computes
+    // ev.treeStack_.back().size() - csgWrapStack.back().groupStartSize
+    // (how many CSGNodes THIS one top-level statement just contributed,
+    // however many that turned out to be -- 0 for a statement whose own
+    // evaluation spliced nothing, e.g. a no-op unknown-module warning; >1
+    // for an attachable()-style multi-body operand) and appends it, as a
+    // Value, onto csgWrapStack.back().groupSizes -- exactly one entry per
+    // top-level geometry statement, in source order, mirroring resolveCsg's
+    // own `groupSizes.push_back(Value{...})` loop.
+    CsgGroupEnd,
+
+    // Closes the matching Op::PushCsgWrap bracket: pops
+    // VmFrame::csgWrapStack's own top entry FIRST (before anything that can
+    // itself throw, e.g. setTreeDepthOrThrow below -- same "exception-
+    // teardown's own pending count must already be right" reasoning as
+    // PopBuiltinWrap), pops the ctx PushCsgWrap unconditionally pushed,
+    // pops ev.treeStack_ to retrieve every group's own CSGNode(s) (flat,
+    // exactly like resolveCsg's own `children` result -- group boundaries
+    // live only in group_sizes, never in the CSGNode list's own shape), and
+    // builds the tagged CSGNode exactly like Evaluator::buildTreeNode's own
+    // post-resolveBody() half does, with params = {"op": site.op,
+    // "group_sizes": ValueList(pending.groupSizes)} -- byte-for-byte what
+    // native resolveCsg returns -- pushing the result onto the new top of
+    // treeStack_. a = index into CompiledChunk::csgWrapSites (same site
+    // Push used).
+    //
+    // teardownVmCallStackDownTo's own exception path (bytecode_vm.cpp) pops
+    // frame->csgWrapStack.size() additional treeStack_ entries per
+    // torn-down frame, alongside its existing builtinWrapStack/
+    // ownsModuleSplice accounting -- same reasoning as PopBuiltinWrap's own
+    // doc comment: an exception can leave N of these brackets open in one
+    // frame.
+    PopCsgWrap,
+
+    // A `children()` / `children(N)` statement -- the runtime-varying
+    // sibling of Op::CallModule/Op::PushBuiltinWrap, closing the LAST
+    // dominant native-reentry source (BOSL2's attachable() calls
+    // children() at nearly every wrapper level; measured 85 of 93 native
+    // reentries in a real script). Unlike PushBuiltinWrap's constructs,
+    // the forwarded children aren't known at compile time (they're the
+    // CALLER's own call-site statements, carried on ctx.childrenNodes/
+    // childrenCallerCtx -- see Evaluator::buildModuleChildCtx), so they
+    // can't compile inline; instead the handler resolves the list at
+    // RUNTIME, looks up/compiles its chunk (the same
+    // childrenListChunkCache_ tryRunCompiledChildren uses, via
+    // lookupOrCompileChildrenListChunk -- gated on useBytecodeVm() &&
+    // inResolvePass_, the cache is pass-scoped), and pushes it onto
+    // vmCallStack_ directly, mirroring Op::CallModule's own zero-native-
+    // call push -- but with a THIRD frame shape: splice-owning like a
+    // module frame (ownsModuleSplice=true, mirroring evalModularCall's
+    // own unconditional splice branch for "children"), yet bracketless
+    // like a bare frame (children() never gets a callStack_/profiling
+    // entry natively either -- only enterUserCall pushes those, and
+    // resolveChildren/builtinChildren never call it). driveVm's existing
+    // completion branch and teardownVmCallStackDownTo both already
+    // handle that combination (their bracket and splice concerns are
+    // independent).
+    //
+    // Handler ordering is load-bearing: checkDebug fires in-handler
+    // against the scope-wrapped ctx (byte-for-byte what Op::
+    // NativeStatement does -- NOT via an emitted Op::CheckDebugStatement,
+    // whose handler passes the un-wrapped ctx); randsBefore is captured
+    // BEFORE argument resolution (rands-in-args taint, same lesson
+    // PushBuiltinWrap already encodes); and treeStack_ is pushed LAST,
+    // immediately before the frame push / native fallback (an early-out
+    // or a throw during arg resolution then has nothing to clean up --
+    // the native path's own catch{pop;throw} has no equivalent here, so
+    // the reorder IS the exception-safety mechanism). The not-eligible
+    // fallback reuses the ALREADY-resolved args (never evalStatement,
+    // which would re-resolve them: double rands(), double side effects)
+    // by inlining evalModularCall's own children branch around a native
+    // evalChildren call. a = index into CompiledChunk::nativeStatements
+    // (the ModularCall node -- arguments, error position, splice node;
+    // no separate site table needed).
+    CallChildren,
+
+    // A single "native passthrough" statement -- intersection_for, every
+    // OTHER builtin module call not covered by Op::PushBuiltinWrap/
+    // Op::PushCsgWrap (see those ops' own doc comments for exactly which
+    // builtins ARE covered -- cube/sphere/hull/linear_extrude/etc., the
+    // ones that never wrap a recursive call in idiomatic OpenSCAD), the `*`
     // modifier's own no-op case, or a user-module call that didn't
     // resolve at compile time (shadowed, forward-declared, or otherwise
     // not statically known) -- anything compileStatementList doesn't give
@@ -344,17 +482,22 @@ enum class Op {
     // pattern that made this op's own native reentry a genuine Windows
     // crash risk in practice, not just a missed optimization. See
     // Op::PushBuiltinWrap's own doc comment for the real story and the
-    // fix.)
+    // fix. children() fell here too, and was the LAST and largest such
+    // reentry source once PushBuiltinWrap's own set was covered -- it now
+    // has Op::CallChildren, above. union()/difference()/intersection()
+    // fell here too, and were the last remaining REAL native-reentry risk
+    // (bespoke group_sizes bookkeeping meant they couldn't just reuse
+    // PushBuiltinWrap's own bracket) -- they now have Op::PushCsgWrap,
+    // above.)
     // a = index into CompiledChunk::nativeStatements. Runtime just does
     // what Evaluator::evalChildren's own per-statement loop already does
     // for one node: derive childCtx via ctx.withScope(...), checkDebug,
     // evalStatement. These are still "leaf-shaped" for what's left here
-    // after Op::PushBuiltinWrap peeled off the proven-risky subset: real
-    // recursion safety for a recursive module chain is covered by
-    // CallModule/PushBuiltinWrap/ForIterNext/the Jump-based if/for control
-    // flow, not by how many of THESE sit alongside them in the same body
-    // -- true again now that the one construct that violated it
-    // (translate()-wrapped recursion) has its own real bytecode instead.
+    // now that Op::PushBuiltinWrap/Op::PushCsgWrap have peeled off every
+    // proven-risky construct: real recursion safety for a recursive module
+    // chain is covered by CallModule/PushBuiltinWrap/PushCsgWrap/
+    // ForIterNext/the Jump-based if/for control flow, not by how many of
+    // THESE sit alongside them in the same body.
     NativeStatement,
 
     // If/if-else's own condition, evaluated NATIVELY (Evaluator::
@@ -690,6 +833,19 @@ struct CompiledChunk {
         const oscad::ASTNode* node = nullptr;
     };
 
+    // One Op::PushCsgWrap/PopCsgWrap site pair -- see those ops' own doc
+    // comments for the full "why union/difference/intersection need
+    // bespoke group_sizes bookkeeping instead of just reusing
+    // BuiltinWrapSite" rationale. `op` is always "union"/"difference"/
+    // "intersection" (the only 3 names resolveDispatch() maps to
+    // resolveCsg) -- always a genuine ModularCall (unlike BuiltinWrapSite's
+    // Modifier kind, this construct has no non-ModularCall variant), so
+    // `node` is typed precisely rather than a generic ASTNode*.
+    struct CsgWrapSite {
+        std::string op;
+        const oscad::ModularCall* node = nullptr;
+    };
+
     // One Op::AssertStatement site -- see that op's own doc comment for
     // the full contract. `conditionArgIndex`/`messageArgIndex` are indices
     // into the site's own argCount-sized popped-argument array (source
@@ -752,6 +908,7 @@ struct CompiledChunk {
     // own doc comments, above) -- always empty for a function chunk.
     std::vector<ModuleCallSite> moduleCallSites;
     std::vector<BuiltinWrapSite> builtinWrapSites;
+    std::vector<CsgWrapSite> csgWrapSites;
     std::vector<AssertSite> assertSites;
     std::vector<const oscad::Expression*> nativeExprs;
     std::vector<const oscad::ASTNode*> nativeStatements;

{openscad_cpp_evaluator-0.13.3 → openscad_cpp_evaluator-0.13.5}/include/openscad_cpp_evaluator/bytecode_vm.hpp

@@ -30,6 +30,24 @@ struct PendingBuiltinWrap {
     int siteIdx = -1;
 };
 
+// One still-open Op::PushCsgWrap bracket's own state -- see that op's own
+// doc comment (bytecode.hpp) for the full contract. `groupSizes` is built
+// up incrementally, one entry per Op::CsgGroupStart/CsgGroupEnd pair (one
+// per top-level GEOMETRY child statement); `groupStartSize` is scratch
+// space for the CURRENTLY OPEN group only (set by CsgGroupStart, consumed
+// by the matching CsgGroupEnd) -- safe as a single scalar, not a stack of
+// its own, because groups within one CSG wrap are siblings in sequence,
+// never nested (unlike PushCsgWrap brackets themselves, which CAN nest,
+// e.g. `union() { difference() {...} }` -- that's what makes
+// VmFrame::csgWrapStack itself a real LIFO, below).
+struct PendingCsgWrap {
+    std::string op;
+    std::uint64_t randsBefore = 0;
+    int siteIdx = -1;
+    std::vector<Value> groupSizes;
+    size_t groupStartSize = 0;
+};
+
 // One ListCompFor/statement-for assignment's own materialized iteration
 // state -- see Op::IterMaterialize/IterReset/IterNext's own doc comments
 // (bytecode.hpp). Lives in the header (not bytecode_vm.cpp's own anonymous
@@ -127,6 +145,15 @@ struct VmFrame {
     // existing invariant ("whoever pops this frame drains its own open
     // brackets first") already covers it.
     std::vector<PendingBuiltinWrap> builtinWrapStack;
+    // Op::PushCsgWrap's own per-frame LIFO -- same role/lifetime/teardown
+    // discipline as builtinWrapStack, just for union()/difference()/
+    // intersection() (see PendingCsgWrap's own doc comment, above, and
+    // Op::PushCsgWrap's, bytecode.hpp). Not explicitly cleared in
+    // releaseVmFrame, same as builtinWrapStack/accumStack/ctxChain -- the
+    // existing invariant ("whoever pops this frame drains its own open
+    // brackets first", normally via matched Push/Pop, or via
+    // teardownVmCallStackDownTo on the exception path) already covers it.
+    std::vector<PendingCsgWrap> csgWrapStack;
     // The ORIGINAL callee name at push time, used by
     // Evaluator::exitUserCallSuccess's own returnHook call when this frame
     // carries a bracket -- deliberately NOT updated by a later tail hop

{openscad_cpp_evaluator-0.13.3 → openscad_cpp_evaluator-0.13.5}/include/openscad_cpp_evaluator/evaluator.hpp

@@ -271,6 +271,25 @@ public:
     // _builtin_children/_eval_children_lazy.
     void builtinChildren(const CallArgs& args, EvalContext& ctx);
 
+    // The "which nodes, evaluated against what context" half of
+    // builtinChildren, shared with Op::CallChildren's runtime handler
+    // (bytecode_vm.cpp) so the native and compiled paths cannot drift on
+    // any of the subtle parts: the caller-ctx re-derivation, the
+    // $-forwarding loop (which must read the post-resolveCallArgs effCtx
+    // -- a children($fn=12) named-$ override lives only there), and the
+    // children(N) statement-index filtering/bounds-check. nullopt means
+    // "nothing to evaluate" (no forwarded children in scope, empty list,
+    // or index out of range) -- a silent no-op in every existing caller,
+    // exactly matching builtinChildren's own early returns. `ctx` must be
+    // the effCtx resolveCallArgs returned, same as builtinChildren's own
+    // parameter today. Public for the same free-function reasoning as
+    // builtinChildren itself.
+    struct ChildrenForward {
+        EvalContext evalCtx;
+        std::vector<const oscad::ASTNode*> nodes;
+    };
+    std::optional<ChildrenForward> prepareChildrenForward(const CallArgs& args, EvalContext& ctx);
+
     // "WARNING: {message}{locSuffix(position)}" via echoFn_, no-op if unset.
     // Public: builtins/import.cpp's not-manifold warning is emitted from a
     // free function, same reasoning as tagGenerated()/builtinChildren().
@@ -532,6 +551,17 @@ private:
     // into the caller's scope exactly the way the native per-statement
     // loop's own writes already do.
     bool tryRunCompiledChildren(const std::vector<const oscad::ASTNode*>& children, EvalContext& ctx);
+public:
+    // The cache-lookup half of tryRunCompiledChildren, shared with
+    // Op::CallChildren's runtime handler (bytecode_vm.cpp, a free
+    // function -- public for the same no-friend-declaration reasoning as
+    // vmCallStack_/treeStack_). Returns the eligible compiled chunk for
+    // `children` or nullptr. Caller owns the useBytecodeVm()/
+    // inResolvePass_ gate -- the pass gate is load-bearing
+    // (childrenListChunkCache_ is pass-scoped, see its own doc comment),
+    // not defensive.
+    const CompiledChunk* lookupOrCompileChildrenListChunk(const std::vector<const oscad::ASTNode*>& children);
+private:
     void evalModularCall(const oscad::ModularCall& node, EvalContext& ctx);
     void evalFor(const oscad::ModularFor& node, EvalContext& ctx);
     void evalLetBlock(const oscad::ModularLet& node, EvalContext& ctx);
@@ -951,15 +981,38 @@ private:
     // body -- so a resolvable module call anywhere in that list gets
     // Op::CallModule bytecode, and any if/for control flow around it gets
     // real Jump-based bytecode too, instead of falling through to the
-    // native per-statement loop. Keyed by the list's own FIRST element,
-    // same convention as assignBlockChunkCache_ (stable and unique per
-    // list: a given evalChildren() call site is always handed the same
-    // leading element across repeated evaluations). Same nullopt-means-
+    // native per-statement loop.
+    //
+    // Keyed by (FIRST element, list SIZE) -- NOT the first element alone,
+    // the original convention borrowed from assignBlockChunkCache_
+    // (where it IS sufficient: a leading-assignment run's first member
+    // uniquely determines the run). Here it wasn't: children()'s
+    // forwarding produces two DIFFERENT lists sharing a first element --
+    // bare `children()` forwards the caller's whole list, `children(0)`
+    // a single-element slice of it -- and keying by front alone made
+    // whichever form ran first poison the cache for the other (caught
+    // for real: `module m() { children(); children(0); }\n
+    // m() { cube(); sphere(); }` made children(0) emit BOTH shapes,
+    // or bare children() emit only one, depending on statement order).
+    // (front, size) fully disambiguates every real producer: fixed AST
+    // lists have unique fronts per call site, and the only same-front
+    // pair (bare vs indexed forwarding) always differs in size except
+    // when the lists are literally identical anyway. Same nullopt-means-
     // tried-and-failed convention, same dangling-pointer hazard, same
     // two-part fix (cleared alongside stmtExprChunkCache_ at the top of
     // every resolveTreeImpl call, only read/written while inResolvePass_)
     // as stmtExprChunkCache_ itself.
-    std::unordered_map<const oscad::ASTNode*, std::optional<CompiledChunk>> childrenListChunkCache_;
+    struct ChildrenListKeyHash {
+        size_t operator()(const std::pair<const oscad::ASTNode*, size_t>& k) const {
+            // Standard hash-combine (boost-style golden-ratio mix) --
+            // either half alone collides by construction here.
+            const size_t h1 = std::hash<const oscad::ASTNode*>{}(k.first);
+            const size_t h2 = std::hash<size_t>{}(k.second);
+            return h1 ^ (h2 + 0x9e3779b97f4a7c15ULL + (h1 << 6) + (h1 >> 2));
+        }
+    };
+    std::unordered_map<std::pair<const oscad::ASTNode*, size_t>, std::optional<CompiledChunk>, ChildrenListKeyHash>
+        childrenListChunkCache_;
 
     // See stmtExprChunkCache_'s own doc comment, immediately above, for why
     // this exists. Not a reentrancy guard (resolveTreeImpl is never called
@@ -1115,6 +1168,17 @@ public:
         return bytecodeVmEnabled() && (!debugHooks_.debugHook || fastContinueBreakpoints_.has_value());
     }
 
+    // Read accessor for inResolvePass_ (private, below) -- Op::CallChildren's
+    // runtime handler (bytecode_vm.cpp, a free function) must gate its
+    // childrenListChunkCache_ access on it, exactly like
+    // tryRunCompiledChildren's own self-gate: the cache is pass-scoped
+    // (cleared per resolveTreeImpl, AST-address-reuse hazard -- see
+    // stmtExprChunkCache_'s own doc comment), and driveVm CAN run outside
+    // the resolve pass (a host calling evalChildren directly reaches
+    // module opcodes via lookupOrCompileModuleChunk, which is NOT
+    // pass-scoped).
+    bool inResolvePass() const { return inResolvePass_; }
+
 private:
     // The fine-grained half of the check above: even when useBytecodeVm()
     // says compiling/using bytecode is on the table at all, a SPECIFIC

{openscad_cpp_evaluator-0.13.3 → openscad_cpp_evaluator-0.13.5}/include/openscad_cpp_evaluator/scope_trail.hpp

@@ -368,12 +368,30 @@ public:
         dirty_[level].push_back(id);
     }
 
+    // Level-aware, NOT a blind pop_back() -- the exact fix
+    // ScopeTrailStorage::popLevel (above) already carries, for the exact
+    // out-of-order-pop bug class its own doc comment describes; this
+    // indexed twin never got the same fix because nothing violated LIFO
+    // view destruction on the dyn trail until Op::CallChildren's
+    // forwarding frame (bytecode_vm.cpp): its evalCtx (a dyn level opened
+    // AFTER the call's own effCtx level) is moved into a VmFrame and
+    // OUTLIVES effCtx, so effCtx's pop runs while evalCtx's later entry
+    // is still physically on top of the same name's stack -- the blind
+    // pop_back() silently removed the still-live forwarded value instead
+    // of the one actually being popped (caught for real: a
+    // children($fn=9) named-$ override read back as the root default 0
+    // inside the forwarded child).
     void popLevel(int level) {
         auto it = dirty_.find(level);
         if (it != dirty_.end()) {
             for (int id : it->second) {
                 auto& stack = stacks_[static_cast<size_t>(id)];
-                if (!stack.empty()) stack.pop_back();
+                for (auto rit = stack.rbegin(); rit != stack.rend(); ++rit) {
+                    if (rit->level == level) {
+                        stack.erase(std::next(rit).base());
+                        break;
+                    }
+                }
             }
             dirty_.erase(it);
         }

{openscad_cpp_evaluator-0.13.3 → openscad_cpp_evaluator-0.13.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "openscad_cpp_evaluator"
-version = "0.13.3"
+version = "0.13.5"
 description = "C++ OpenSCAD evaluator with Python bindings"
 readme = "README.md"
 requires-python = ">=3.12"

{openscad_cpp_evaluator-0.13.3 → openscad_cpp_evaluator-0.13.5}/src/bytecode_compiler.cpp

@@ -1202,6 +1202,45 @@ public:
         out.push_back({Op::PopBuiltinWrap, idx, 0, &wrapperNode.position()});
     }
 
+    // union()/difference()/intersection() -- see Op::PushCsgWrap's own doc
+    // comment (bytecode.hpp) for why these can't just reuse emitBuiltinWrap:
+    // resolveCsg (booleans.cpp) needs per-top-level-child-statement
+    // "group_sizes" bookkeeping, replicated here as one Op::CsgGroupStart/
+    // CsgGroupEnd pair per GEOMETRY child, with every ASSIGNMENT child
+    // compiled first, unconditionally, regardless of interleaving in
+    // source -- mirrors resolveCsg's own two-pass split (`assignNodes`
+    // fully evaluated, THEN one evalChildren call per `geoNodes` entry)
+    // exactly, including its ModuleDeclaration/FunctionDeclaration
+    // exclusion (a nested declaration inside a CSG block contributes to
+    // neither pass -- already hoisted into scope, nothing to run here).
+    void emitCsgWrap(const oscad::ModularCall& call, std::vector<Instruction>& out) {
+        out.push_back({Op::CheckDebugStatement, internNativeStatement(&call), 0, nullptr});
+        CompiledChunk::CsgWrapSite site;
+        site.op = call.name->name;
+        site.node = &call;
+        chunk_.csgWrapSites.push_back(std::move(site));
+        const int idx = static_cast<int>(chunk_.csgWrapSites.size()) - 1;
+        out.push_back({Op::PushCsgWrap, idx, 0, &call.position()});
+
+        std::vector<const oscad::ASTNode*> assignNodes;
+        std::vector<const oscad::ASTNode*> geoNodes;
+        for (const auto& c : call.children) {
+            if (c->kind() == oscad::NodeKind::Assignment) {
+                assignNodes.push_back(c.get());
+            } else if (c->kind() != oscad::NodeKind::ModuleDeclaration &&
+                       c->kind() != oscad::NodeKind::FunctionDeclaration) {
+                geoNodes.push_back(c.get());
+            }
+        }
+        compileStatementList(assignNodes, out);
+        for (const oscad::ASTNode* geoNode : geoNodes) {
+            out.push_back({Op::CsgGroupStart, 0, 0, nullptr});
+            compileStatementList(std::vector<const oscad::ASTNode*>{geoNode}, out);
+            out.push_back({Op::CsgGroupEnd, 0, 0, nullptr});
+        }
+        out.push_back({Op::PopCsgWrap, idx, 0, &call.position()});
+    }
+
     void compileOneStatement(const oscad::ASTNode& stmt, std::vector<Instruction>& out) {
         using oscad::NodeKind;
         trackSpan(stmt);
@@ -1420,12 +1459,32 @@ public:
                     emitBuiltinWrap(*wrapKind, call.name->name, call, children, out);
                     return;
                 }
-                // Every other builtin, children(), or a name that didn't
-                // resolve to a user module statically -- native
-                // passthrough, exactly like echo/assert/etc. below. Never
-                // the recursion-depth risk this compiler targets for
-                // THESE (see NativeStatement's own doc comment,
-                // bytecode.hpp).
+                // children() -- the runtime-varying forwarding builtin,
+                // detected by the same function-pointer-identity pattern
+                // as Transform/Color above. Deliberately NO preceding
+                // Op::CheckDebugStatement (unlike emitBuiltinWrap's own
+                // emission): CallChildren's handler fires checkDebug
+                // itself against the SCOPE-WRAPPED ctx, byte-for-byte
+                // matching Op::NativeStatement's own handler -- the
+                // CheckDebugStatement handler passes the un-wrapped ctx,
+                // which would be a subtle behavior change for this node.
+                // See Op::CallChildren's own doc comment (bytecode.hpp).
+                if (dispatchIt != dispatch.end() && dispatchIt->second == &resolveChildren) {
+                    out.push_back({Op::CallChildren, internNativeStatement(&stmt), 0, &stmt.position()});
+                    return;
+                }
+                // union()/difference()/intersection() -- see emitCsgWrap's
+                // own doc comment for why these need their own bespoke
+                // bracket rather than emitBuiltinWrap's.
+                if (dispatchIt != dispatch.end() && dispatchIt->second == &resolveCsg) {
+                    emitCsgWrap(call, out);
+                    return;
+                }
+                // Every other builtin, or a name that didn't resolve to a
+                // user module statically -- native passthrough, exactly
+                // like echo/assert/etc. below. Never the recursion-depth
+                // risk this compiler targets for THESE (see
+                // NativeStatement's own doc comment, bytecode.hpp).
                 out.push_back({Op::NativeStatement, internNativeStatement(&stmt), 0, &stmt.position()});
                 return;
             }