fix(cpp): correct args counting for operator() and constructors with initializer-lists

[squid-protocol]

Summary

Fixes #2012. Two distinct, independently-diagnosed bugs in C++ args counting:

Pattern 1 (zero-undercount): operator() and out-of-class methods with non-whitelisted parameter types (e.g. mlir::ModuleOp) reported args=0. Root cause was two-fold: (a) the regex-based args extraction rejects operator() syntax and enforces a rigid parameter-type whitelist, silently defaulting to 0 on failure; (b) the structural fallback counter needed to skip operator()'s own empty name-parens to find the real parameter-list parens.

Pattern 2 (off-by-one overcount): a zero-arg constructor with a member-initializer-list (Ctor() : member(x) { }) reported args=1 -- the args-search text swept the initializer-list clause into the parameter-list search, misreading member(x) as a parameter.

Fix

• Truncate the args-search text at the first top-level : (excluding ::) before the opening brace, cutting the initializer-list clause out of the search entirely (Pattern 2).
• Fall back to the structural _count_top_level_args counter for c/cpp when the regex path fails, since func_start already validated it's a real function (Pattern 1a).
• Skip operator()'s own empty name-parens in that structural counter to find the real parameter list (Pattern 1b).

Both fixes gated to lang_id/primary_lang_id in ("c", "cpp"). The common case (real params + initializer list) is unaffected.

Evidence

Case	Before	After
operator()(a, b)	0	2
Out-of-class BuildBuffer(mlir::ModuleOp module)	0	1
Zero-arg ctor with member-initializer-list	1	0
Ctor with real params + initializer-list (common case)	2	2 (unchanged)

Corpus-wide: args_exact_match improved from 1233 to 1295 (out of 1296/1297 comparable functions).

Known minor limitation

The operator() detection uses an 8-character trailing substring match (args_str[open_idx-8:open_idx] == "operator") without a word-boundary check. An identifier ending in ...operator( with no separator before it (e.g. a hypothetical custom_operator(...)) could theoretically false-positive and skip a real first parameter list. No such case was found in the ~80-repo corpus; flagging for awareness rather than blocking on it.

Verification

• pytest tests/core_engine/test_detector.py tests/extraction/languages/test_cpp.py -q -- 195 passed (1 new regression test covering both patterns + the unaffected common case)
• audit_check.py --ci -- clean
• crucible_check.py --mode both -- drift confined to cpp (NVDA/godot/mlir Function Analysis changes, attributable) plus expected global mass/coordinate ripple in unrelated directory groups (confirmed no Function Analysis changes outside cpp), blessed
• tree_sitter_accuracy_audit.py --all --ci -- clean
• tri_comparison_chart.py --all --ci -- clean

This touches shared detector.py helpers (_calculate_block_metrics, _count_top_level_args), so the full --all --ci sweep was run per repo convention.

🤖 Generated with Gemini (via agy), orchestrated and independently verified by Claude Code.