feat(cpp-phase2): PR-04 resolution-quality fixes (glibc attrs, transitive includes)

[shivasurya]

Summary

PR-04 of the C/C++ Phase 2 stdlib stack — three resolution-quality fixes uncovered during local validation against redis and proxygen. Stacked on PR-03; should merge after.

What's in here

Three independent fixes, each in its own commit:

1. Recover glibc-decorated declarations (fix(generator))

Tree-sitter's C grammar can't parse:

extern size_t strlen (const char *__s)
     __THROW __attribute_pure__ __nonnull ((1));

It collapses the whole declaration into an ERROR node and the symbol vanishes. Effect: strlen, strcmp, strncmp, strcasecmp, memcmp, memmem, snprintf, vsnprintf and dozens of similar glibc functions never made it into the manifest.

tools/internal/clikeextract/clike_preprocess.go strips the macros to whitespace before tree-sitter parses (length-preserving, line-number-accurate). Macro list sourced from <sys/cdefs.h>. Wired into both C and C++ extractors.

2. Resolver tolerant of generator's map placement (fix(registry))

std::move is emitted into manifest's functions map but PR-02's GetFreeFunction only consulted free_functions. Lookup now falls through to both maps. Strict superset of previous behaviour.

3. Transitive-include fallback (feat(builder))

Real-world C++ relies heavily on transitive includes. A file that calls std::move without directly #include <utility> (utility comes through <vector> etc.) used to fail resolution. The resolver now walks every manifest header as a fallback, bounded to qualified names so unqualified project symbols can't accidentally bind. Required adding ListHeaders() to the CStdlibLoader / CppStdlibLoader interfaces.

Validation results

Resolution-report on real codebases, regenerated registries, file:// loader.

Project	PR-03 baseline	PR-04
smoke fixture (testdata/c/stdlib)	71% (5/7)	100% (7/7)
redis	60.6%	63.1% (+2.5 pp)
proxygen	12.0%	17.1% (+5.1 pp)

Top-1 unresolved in proxygen (std::move, 1603 occurrences) is gone — all std::move calls now resolve via the transitive fallback. Total Linux C functions extracted: 8,467 → 9,205 (+738).

The proxygen number is still well below the 98% spec target, but the remaining unresolveds are largely non-stdlib: gtest / glog / folly logging macros (VLOG, EXPECT_EQ, CHECK), folly types (folly::makeUnexpected, hasError), and unqualified class-method calls that need receiver-type inference (separate concern, future PR).

Verification

• gradle buildGo — clean
• go test ./... — all packages pass
• golangci-lint run ./... — 0 issues

Out of scope (future PRs)

• Receiver-type inference for chained STL method calls (vec.size(), str.empty())
• gtest / glog / folly macro recognition (project-specific overlay)
• Generator emitter cleanup so namespaced symbols always land in free_functions

Test plan

• CI green
• Operator validation: regenerate Linux registries from this branch and re-run resolution-report against own codebases

🤖 Generated with Claude Code

[safedep]

SafeDep Report Summary

No dependency changes detected. Nothing to scan.

View complete scan results →

_{This report is generated by SafeDep Github App}

[github-actions]

Code Pathfinder Security Scan

No security issues detected.

Metric	Value
Files Scanned	14
Rules	205

---

_{Powered by Code Pathfinder}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.68%. Comparing base (2d5c149) to head (f0bfd81).

Additional details and impacted files

@@                       Coverage Diff                        @@
##           shiva/cpp-phase2-pr03-remote     #682      +/-   ##
================================================================
+ Coverage                         85.67%   85.68%   +0.01%     
================================================================
  Files                               202      203       +1     
  Lines                             29141    29209      +68     
================================================================
+ Hits                              24966    25029      +63     
- Misses                             3203     3207       +4     
- Partials                            972      973       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[shivasurya]

Pure-stdlib gap analysis (third-party / macros excluded)

Cross-referencing each unresolved call's target name against the actual symbols in our linux/c and linux/cpp manifests reveals two distinct unresolved buckets — both fully fixable on the stdlib side, neither addressed by this PR but both worth queueing as follow-ups.

C — transitive-include fallback gap

3,107 redis unresolved calls have targets that already exist in our linux/c manifest. Top:

Symbol	Count
strcasecmp	952
memcpy	146
strlen	136
snprintf	128
strerror	119
memset	114
exit	104
free / malloc	89 + 63
strcmp	82
printf	81
fclose / fopen	42 + 20
pthread_mutex_lock / unlock	27 + 27
ntohs / htons	30 + 17

Root cause: redis files #include "server.h", and server.h pulls in <string.h>, <stdio.h>, <stdlib.h>, <pthread.h>, etc. The PR-02 C resolver walks only the caller file's direct system includes, so when redis-cli.c calls strlen it never looks in <string.h> because that header is brought in transitively through server.h.

Fix: mechanical port of this PR's C++ transitive fallback (lookupCppStdlibFreeFunction) to the C resolver's lookupCStdlib. About 30 LoC. Bounded to the manifest's header set so no risk of binding project symbols.

Estimated uplift: redis 63.1% → ~67.5% (+4.4 pp). Similar shape on any C codebase that uses an internal "common.h" pattern, which is essentially every non-trivial C project.

C++ — receiver-type inference gap

~3,621 proxygen unresolved calls have targets that exist as STL class methods. Top:

Method	Calls	Likely class
empty	525	vector / string / map / optional
size	460	vector / string / map
append	221	string
end	206	iterator endpoints
emplace_back	202	vector
value	166	optional
reset	154	unique_ptr / shared_ptr
find	145	string / map / set
front / back	137 + 54	vector / deque
data	136	string / vector
push_back	132	vector
begin	122	iterator endpoints
length	105	string
erase	93	container methods
has_value	89	optional
release	85	unique_ptr
clear / reserve	66 + 64	container methods
c_str	64	string

Root cause: these are unqualified vec.size() / str.empty() / opt.value() calls. The resolver needs to know the receiver's type to dispatch through GetMethod("vector", "std::vector", "size"). Today the type engine doesn't propagate types through:

• function returns (auto v = obj.getVec();)
• class members (vec_.size() inside a method when vec_ is a private field)
• chained calls (obj.getVec().size())

Fix: larger work — extending the type engine's propagation rules. Estimated uplift: proxygen 17.1% → ~30% (+13 pp).

Out of scope (third-party / macros)

• gtest/glog macros (VLOG, EXPECT_EQ, CHECK, WillRepeatedly) — ~5,500 proxygen calls
• folly library (folly::makeUnexpected, folly::to, chainLength) — ~700
• redis project API (RedisModule_*, serverLog, REDISMODULE_NOT_USED) — thousands

Recommended sequence

1. PR-05 — port the transitive-include fallback to C. Small, mechanical, ships the redis +4.4 pp uplift.
2. (Future) — Receiver-type inference. Larger, ships the proxygen +13 pp uplift.

Both target the stdlib-resolved-rate metric specifically, not the noisier "% of all calls" overall number.