test(cpp): C++ special-member name reconciliation → function precision 100%

[squid-protocol]

tree_sitter_accuracy_audit measured cpp function precision at 95.0% (68 "extra"
functions). Investigating every one: they are all real C++ special members that GitGalaxy
correctly extracts and tree-sitter-cpp either names differently or drops. Zero genuine
GitGalaxy false positives — GitGalaxy's cpp function precision is genuinely 100%; the metric
was under-counting it.

Three shared-reader fixes

(all in tree_sitter_accuracy_audit.py, used by both the accuracy audit and
tri_comparison_gatherer)

1. Conversion-operator naming

tree-sitter's operator_cast / qualified_identifier node text bakes in the parameter list and
cv-qualifiers — Variant::operator String() const — and, for a template conversion, the
argument list — operator Vector<Plane>() const. GitGalaxy's func_start stores just
Variant::operator String / operator Vector. Added _cpp_canonical_operator_name: strips the
trailing (...) + cv-quals and <...> args, keeps operator and the Ns:: qualifier.
Guarded to conversion operators only (operator + a type token) — symbol operators
(operator(), operator==, operator[], operator->) are left untouched.

2. _FORCE_INLINE_-macro parse errors

A function-like macro with no visible definition immediately before a special member desyncs
tree-sitter-cpp: it drops the operator keyword or the ~ into a sibling ERROR node and
names the member by its bare type (_FORCE_INLINE_ operator T() const → "T";
_FORCE_INLINE_ ~Variant() → "Variant"). Recover the prefix from the ERROR node so the two
readers agree instead of double-counting.

3. Blind-spot promotion

Added cpp to _find_blind_spot_ranges (small tree-sitter-cpp ERROR nodes, padded ±2 lines for
the template <...> header line GitGalaxy anchors on), so GitGalaxy's correct extraction of a
constructor tree-sitter's error recovery loses (_FORCE_INLINE_ RequiredResult(...) : ... {})
is promoted to ground truth via the existing #1849 Phase 2 mechanism.

Result

Metric	Before	After
Func precision	95.0%	100.0%
extra_functions	68	0
Func recall	87.0%	89.3%
Args exact-match	99.8%	100.0%

tree_sitter_accuracy_audit --all --ci 31/31 OK; C unchanged. cpp baseline + summary table
regenerated.

Tri-comparison

The shared _get_node_name fix also lands in tri_comparison_gatherer, moving ~55
conversion-operator occurrences from cpp/function/existence/agree[tree_sitter]_vs[ctags,gitgalaxy]
(200 → 136) and agree[ctags,gitgalaxy]_vs[tree_sitter] (66 → 14) into GitGalaxy+tree-sitter
consensus. Verdicts, chart, points-of-interest regenerated; every reproducing ledger shape stays
validated. tri_comparison_chart --all --ci OK.

Scope

No engine change — gitgalaxy/core/* untouched, no golden-master re-bless. This is a
measurement-accuracy fix: the audit now correctly recognizes that GitGalaxy and tree-sitter
found the same C++ members.

🤖 Generated with Claude Code

https://claude.ai/code/session_019Zm1uVBpVEZJ9SB7bWFR8S

[github-actions]

🐦‍⬛ Muninn Security Scan

✅ No security issues found.

🐦‍⬛ Powered by Muninn · Skald Lab