codegraph/docs/design/ccpp-kernel-port-checklist.md

C/C++ kernel port (R7a) — the bug-for-bug checklist

Status: COMPLETE — walker SHIPPED + gates PASSED, c/cpp DEFAULT-ROUTED (2026-07-17). Walker: codegraph-kernel/src/ccpp/mod.rs (one dual-language module, every branch below mirrored; its header comment lists the quirks). Grammars: tree-sitter-c v0.24.2 (b780e47, parser.c f2883ff9…) + tree-sitter-cpp v0.23.4 (f41e1a0, parser.c 2a35a43b…, scanner.c cf60387d…), crates pinned =exact in Cargo.toml, wasm vendored from the same tags, kernel-grammar-parity green. preParse HOISTED to the route point (preParsedSource in src/extraction/kernel/index.ts — both tryKernelExtract and the raw bulk path), so no blanking ported to Rust.

Gate results (2026-07-17):

• Parity sweeps — 0 diffs on every compared file: redis 592, git 790, fmt 42, protobuf 925, ALS-Community (UE spot-check) 40 files byte-parity.

• Full-init dump-diffs — byte-identical kernel-arm vs wasm-arm: redis (131,921 dump lines), git (160,844), fmt (35,433), protobuf (780,620), ALS (3,694).

• Torture fixtures torture.c/.cpp/.hpp + CRLF variants + Metal/CUDA hoist-parity + defer tests in __tests__/kernel-ccpp-parity.test.ts; new preParse blanks unit-tested in extraction.test.ts; full suite green with CODEGRAPH_KERNEL_EXPECT=1.

• Deferral-rate guard — CORRECTED BY MEASUREMENT (the §4f pattern): the <10% bar was calibrated on ts/java/py/go (0–0.42% parse-error incidence). Macro-heavy C/C++ genuinely parses with errors at double-digit file rates (final sweeps: als 9%, git 16.1%, redis 25.3%, protobuf 25.8%, fmt 42% — fmt's template metaprogramming + .operator[]-in-decltype shapes are grammar-inherent), and every erroring file defers BY POLICY. Measured with the defer disabled (CODEGRAPH_KERNEL_CCPP_ERROR_EXTRACT=1, sweep-only hatch): recovery-divergence is real (21/207 redis, 8/382 git, 9/31 fmt erroring files extract differently across UTF-8/UTF-16), so the defer stays. The sweep harness now takes --max-deferral (default 0.1; use 0.5 for c/cpp — a broken walker still trips it by deferring ~everything).

• Seven NEW/extended preParse blanks cut real incidence (from 32%/52% starting points; linux kernel/+mm/ subtrees 79% → 58%), each offset-preserving, TS-side, shared by both arms, and a graph-quality win for the wasm path itself (git 7.1k → 13.3k nodes; the linux subtrees 2.4k → 7.2k — NOTE these are parity-sweep compared-file totals, NOT full-graph counts: round 2 established that full-graph node deltas are small because wasm error recovery was already salvaging most SYMBOLS on deferred files; the blanks' full-graph win is EDGES + phantom cleanup — see the round-2 record below): #ifdef __cplusplus guard bodies, lone macro lines (FMT_BEGIN_NAMESPACE, Q_OBJECT), C statement iterator macros (list_for_each_entry(…) { } and brace-less bodies), C trailing param attrs (int argc UNUSED), the curated Linux/sparse annotation list (__init/__user/… — structural matching is impossible there: __u32 count is shape-identical; parameterized __printf(1,2) guarded out) + container_of's type-keyword argument, leading attr macros extended to cpp, and directive-line restore (stops the older blanks corrupting #define lines).

• Defer-reuse (the linux-economics fix): a deferred file used to pay the pipeline three times — the worker's raw kernel try, extractFromSource's kernel RE-try, then wasm with a third preParse. A one-slot defer memo in the route point short-circuits the repeat kernel attempt and hands the already-blanked source to the wasm fallback (sourceIsPreParsed). On linux this + the annotation blanks took the kernel-arm parse-loop from 560s (WORSE than the 426s wasm arm) to 356s vs 435s wasm-arm (−18%), and the 2c/6GB envelope to 19.1 min kernel-arm vs 22.9 min wasm-arm (−17%) with a RICHER graph (2,048,295 nodes / 6,406,933 edges; two independent kernel-arm runs byte-same counts).

• Linux-scale dump gate: kernel-arm and wasm-arm full graphs are byte-identical — dump-graph.mjs over both 5.2GB DBs: 10,444,551 dump lines each, sha256 cd4182e6… on both. (Dumping a 2M-node DB needs a big-heap host run — `node --max-old-space-size=16000 …

  file` then hash the FILE; the in-container 6GB heap OOMs and a straight pipe at GB scale dies with ENOBUFS, both of which silently hash truncated output as the empty stream.)

• Deferral round 2 (2026-07-18) — eight new C-only passes + word-list extensions took the linux kernel/+mm/ deferral 58.6% → 33.9% (483 → 279 of 824 files; census-driven: bucket-defers.mjs clusters each deferring file's FIRST post-preParse error-line shape). New passes, all in preParseCSource: parameterized-annotation whole-blank (__free(kfree), __printf(4,0), __counted_by(n), __bpf_md_ptr(…) — extends through a stranded field ;, since a lone-; field errors but an empty struct body doesn't), type-keyword-arg scanner (kzalloc_obj(struct T), list_entry(p, struct T, m) — a bounded hand scanner, NOT a nested regex (backtracking risk on untrusted repos); head exclusions sizeof/offsetof/_Generic/va_arg + call-vs-declaration guard: an identifier or * before the head rejects, return excepted; blanks trailing stars so DEFINE_PER_CPU(struct T *, x) leaves two plain idents), static|extern CAPS_MACRO(…); whole-line blank at ANY scope (bare EXPORT_SYMBOL(x); parses natively as a K&R decl — probed — and stays), the ONE REWRITE in the family — static CAPS(type, name) = { → static <type> <name> = { (name + tail keep exact offsets via d-flag indices; blanking would strand the brace block or discard initializer fn-refs), va_arg(ap, const char *) second-arg blank (single-token va_arg(ap, int) parses natively — probed), GNU named-variadic #define f(args...) DOTS-only blank (post-restoreDirectiveLines by design; whole-tail blanking fails — #define NAME + trailing spaces errors, measured), storage-sandwiched lowercase markers (static notrace void — the sandwich is the guard), C23 auto (auto x = only; auto int x untouched), and multi-line spans for the statement-iterator-macro blank (hlist_for_each_entry_rcu(…,\n lockdep_is_held(&m)) { — bails on ;/braces mid-span). Word list += cacheline family (2- AND 4-underscore spellings), __noclone, __lockfunc, __ref, __private, __bitwise, __nosavedata, __no_kcsan, __cpuidle, __ksym, __net_initdata, __initdata_memblock/_or_meminfo. Cross-repo: git 16.1 → 12.2%, redis 25.3 → 24.1%, fmt/protobuf unchanged (cpp-dominant — C-only passes, correct), 0 diffs all five sweeps. Linux full-tree (2c gate runs): kernel-arm parse 356 → 306s, envelope 19.1 → ~17.1min (host- contaminated, indicative), counts 2,049,153 / 6,413,518 (+858 nodes, +6,585 edges vs R7a — the SYMBOLS were mostly already error-recovered; the graph win is relationships + phantom cleanup). Deliberately skipped: #ifdef CONFIG_X if/else interleaves + labels (genuine preprocessing), TP_PROTO/TRACE_EVENT DSL headers (no real code to recover), module_init(x) without ; (K&R-definition ambiguity), single-token va_arg (native). Torture fixture grew a round-2 section (both-arm parity pinned); unit tests in extraction.test.ts; suite 2517 green.

The sections below are §0a-recipe step 1's output — every TS-side branch the walker mirrors, with file:line anchors (as of 705e501). Read WITH docs/design/rust-kernel-migration-plan.md (§0a recipe, §5 gates).

Architecture decisions (already made by the plan)

1. preParse stays TS-side and is HOISTED to the route point. tryKernelExtract(filePath, source, lang) currently receives RAW source (tree-sitter.ts:6707); the wasm path applies extractor.preParse inside TreeSitterExtractor (tree-sitter.ts:488). For kernel-routed c/cpp, apply the SAME preParse before the kernel call so both arms parse identical blanked bytes — all seven blanking passes (preParseCppSource / preParseCSource, languages/c-cpp.ts:698/750) then need NO Rust port, and every offset survives (they're all equal-length-space replacements).
2. Metal + CUDA ride the cpp route (corrected from the first draft): .metal/.cu map to language 'cpp' at detectLanguage (grammars.ts:135), so there is no separate routing decision — when cpp routes to the kernel, those files come along as blanked cpp. The preParse hoist MUST pass filePath (the extension gates Metal-attribute blanking) and the CUDA content gate rides for free. Their suite tests are the parity insurance.
3. One walker module, dual language (codegraph-kernel/src/ccpp/), flagged c vs cpp like tsjs/ flags its four dialects. Grammars: tree-sitter-c + tree-sitter-cpp crates, wasm vendored from the SAME tags (sha-matched parser.c + scanner, ts-cli 0.25.10). Upgrade the production wasm FIRST and get the full suite green before the walker exists (isolate grammar-bump effects, as R2 did for TS/JS).

Extractor configs (languages/c-cpp.ts — read the whole file when porting)

cExtractor (line 180): functionTypes=[function_definition]; NO class/method/interface types; structTypes=[struct_specifier] (superseded: union_specifier joined structTypes — a named union U { … }; is a definition, extracted with kind struct, and typedef union { … } N; resolves through resolveTypeAliasKind like typedef struct); enumTypes=[enum_specifier]; enumMemberTypes=[enumerator]; typeAliasTypes=[type_definition]; importTypes=[preproc_include]; callTypes=[call_expression]; variableTypes=[declaration]; nameField=declarator; isConst = any child type_qualifier with text const; getReturnType = extractCppReturnType; resolveTypeAliasKind = typedef enum/struct with body → that kind (anon inner specifier takes the typedef's name); extractImport: system_lib_string → strip <>, else string_literal>string_content; signature = full #include line. recoverMangledName = recoverMangledCppName (post-parse salvage, PORT).

cppExtractor (line 755): adds classTypes=[class_specifier] with skipBodilessClass (forward decls / elaborated refs mint no node, #1093); methodTypes=[function_definition] (method vs function decided by isInsideClassLikeNode); typeAliasTypes += alias_declaration (using X = …); resolveName = extractCppQualifiedMethodName (line 75: macro-recovered name first, else LAST :: segment of the declarator's qualified_identifier — found by BFS that SKIPS parameter_list + trailing_return_type, line 13); getReceiverType = extractCppReceiverType (line 86: the qualifier prefix, stripCppTemplateArgs-normalized — multi-line template args must not leak newlines into qualifiedName, #1286/NAME_MAX); getReturnType = extractCppReturnType → normalizeCppReturnType (line 122: smart-ptr/optional unwrap to pointee, cv/template/ptr strip, last :: segment, primitives → none — set CPP_NON_CLASS_RETURN line 108); getVisibility = nearest preceding access_specifier scan in the parent's children (line 785); isMisparsedFunction (line 811): name starts namespace, name ∈ {switch,if,for,while,do,case,return}, or isMacroMisparsedTypeDecl (line 261: bodyless class/struct specifier in type + non-function_declarator declarator → DROP the node).

Name-salvage helpers to port exactly: recoverCppMacroDefinedName (line 49 — ALL-CAPS-with-underscore parsed name + first param a LONE lowercase type_identifier + ≥2 params + NO other lone-identifier param; gtest TEST_F(Fixture, Name) / PYBIND11_MODULE(ext, m) bail); recoverMangledCppName (line 406 — only already-mangled names, Ret (name) idiom left alone, last token before (, primitive/keyword guard); stripCppTemplateArgs (line 157 — depth-counted removal of every balanced <…>); cDeclaratorIdentifier (tree-sitter.ts:234 — declarator chain walk, function_declarator → null, 12-hop guard).

tree-sitter.ts branches (anchors as of 705e501)

Line	Mechanism	Must-mirror details
962	cpp namespace prefix stack (#1291)	named namespace_definition pushes its name (C++17 a::b as written) onto the QN prefix while walking children; anonymous falls through bare
2795	C file-scope variables	only when NO function ancestor; iterate declarators; accept ONLY init_declarator / pointer_declarator / array_declarator — a BARE identifier declarator is a macro-prototype misparse, skip (loses uninit scalars by design); name via cDeclaratorIdentifier; signature = <first 100 chars>; kind constant/variable via isConst
4313	explicit operator calls (#1247)	callee = function field + ERROR-wrapped operator_name sibling; compact symbolic spacing (operator *→operator*, word forms keep space); receiver ->→.; DROP unless receiver is this (bare name) or simple identifier/member chain (silent miss over wrong edge)
~4340	field_expression method calls	recv.method/ptr->method → recv.method; SKIP_RECEIVERS {self,this,cls,super} → bare name; LITERAL receiver → emit nothing (#1230)
4398	call-result receivers (#645/#608)	receiver is call_expression → <innerCallee>().<method> re-encode (c AND cpp in the gate list)
4534	template-arg strip on callees	callee contains < and NOT operator → stripCppTemplateArgs (CUDA launch sites post-blank take this shape)
4545	local fn-ptr call fan-out (#932-adjacent)	bare-identifier callee found in cppLocalFnPtrs[callerId] → emit one calls ref PER recorded target, suppress the local name
5173	stack construction (#1035)	cpp declaration where type ∈ {type_identifier, template_type, qualified_identifier} and any init_declarator has value ∈ {argument_list, initializer_list} → extractInstantiation
5183	fn-ptr binding recording	inside a body: declaration>init_declarator (identifier declarator) or assignment_expression (identifier left); value must be pointer_expression whose child(0) is &; target ∈ {identifier, template_function, qualified_identifier}, template-stripped; per-callerId map of per-local Sets (branch reassignments accumulate)
5408	base_class_clause → extends (#1043)	per base: type_identifier / qualified_identifier / template_type, stripCppTemplateArgs'd; access-specifier keywords skipped
4740	static member refs — cpp only (c not in STATIC_MEMBER_LANGS, line 345)	Foo::BAR value reads → references edge; VERIFY the MEMBER_ACCESS_TYPES shapes for qualified_identifier + the call-callee skip during the port
—	value-reference edges	c: YES (VALUE_REF_LANGS line 401), cpp: NO. Port the value-ref machinery for C only (crib go.rs / tsjs — shadow prune, scope stack, MAX_VALUE_REF_NODES cap, CODEGRAPH_VALUE_REFS=0 kill)
—	fn-ref capture (#756)	function-ref.ts:376: c: cFamilySpec(), cpp: cFamilySpec({ addressOfOnly: true }); note line ~582: &Cls::m exemption from the bare-ids-are-free-functions rule — read cFamilySpec fully when porting
355	INSTANTIATION_KINDS	includes new_expression → cpp new Foo(...) instantiates (verify the cpp entry list when porting)

Generic paths c/cpp share with ported languages (already mirrored by java.rs/python.rs/go.rs — re-verify each against c/cpp fixtures rather than re-deriving): extractFunction/Method QN via scope stack + receiverType (Recv::name like Go), struct/enum/enum_member extraction, typedef → type_alias (+ resolveTypeAliasKind kind override), imports, docstrings (preceding comment nodes — docstring.rs already handles C markers incl. the CRLF semantics from #1329), contains edges, signature truncation in UTF-16 units, MAX_FILE_SIZE / generated-file skips, has_error() → defer:.

Gates (per plan §5, no exceptions)

• Torture fixtures: torture.c (fn-ptr tables, typedef enum/struct, file-scope consts incl. multi-declarator, macro-prototype misparse shape, value-refs) + torture.cpp/torture.hpp (namespaces incl. C++17 nested, out-of-line Cls::method defs, templates + template bases, operators incl. spaced call sites, stack construction, local fn-ptrs, UE-macro shapes THROUGH the hoisted preParse, using aliases, anonymous namespace, access specifiers).
• Parity sweeps + full-init dump-diffs byte-identical: redis + git (C), fmt + a protobuf-class repo (C++), plus a UE-macro-heavy spot-check (blanking-hoist parity) — then linux in cg1212 (expect parse 338s → ~120–180s at the 2c envelope; graph counts must stay 2,048,664/6,405,964).
• Deferral-rate guard <10%; suite; changelog rides the existing kernel entry.
• DEFAULT_ROUTED += c, cpp only after ALL of the above.