fix(resolution): a definition its language makes file-local is not a cross-file target

[danusha2345]

Fixes #1730. Fixes #1731. Standalone off main (b9ca4b7), two commits — the second narrows the C rule to translation units after the first cut removed 4,306 real calls onto header static inlines; details in the comment below.

What

Name matching accepted any same-named definition as the target of a reference from another file, however the language scopes that definition. One test, isVisibleAcrossFiles, now declines a candidate in another file when:

language	rule	source of truth
C / C++	the function is defined in a source file (.c .cc .cpp .cxx .m .mm) and its definition line carries static — a header's static inline is textually included into every unit that names it and stays visible	read from source (the extractor records no storage class; the kernel arm would need the same field)
Kotlin, Java, C#, Swift, Scala, Dart, PHP	visibility === 'private'	extractor
Go	lowercase identifier and a different directory	the name's case — the extractor's isExported is unset for every Go method, exported or not
Rust	non-pub item and the reference is outside the item's module subtree	extractor + path: src/net.rs / src/net/mod.rs own src/net/; a child reaches its ancestors' private items via super::. A method in an impl Trait for Type block has the trait's visibility and is exempt (read from the enclosing impl header)

Same-file candidates are always visible.

Where the test runs, and why there

In ReferenceResolver, on the target the whole name-matching pipeline settled on (exact-match, fuzzy, instance-method, qualified-name all come through it), so a rejection ends the reference unresolved. The first cut declined inside matchByExactName, #1720-style, and that let the ref fall through to matchFuzzy, which committed to a same-language namesake the ranking had passed over: eight new fuzzy edges on one tree, every one onto a JavaScript fail in a different file, the real fail being a local const fail = (msg) => … the graph does not hold (#1669's shape). matchFuzzy checks its own survivor as well, since nothing runs after it.

Measurement

Five corpora, each indexed at b9ca4b7 and at this branch (wasm arm), edge rows joined back to symbols and keyed with resolvedBy:

corpus	files	LOST	GAINED
betaflight fork (C)	2,109 .c	145	0
Android / Go / Electron app	114 .kt, 42 .go, 96 .js	142	0
emmc-reader-gui (Rust workspace)	71 .rs	195	0
skylab_hub (Rust + Kotlin)	35 .rs, 4 .kt	92	0
vitejs/vite (JS/TS only)	1,635	0	0

Read back from source, the removals are the shapes in the two issues: STM32 USB class sources → static handlers in GD32's usbd_enum.c, usbd_get_descriptor → a static get_device_descriptor in a USB class file it never links; editor.apply() → an unrelated class's private fun apply; latch.await() in tests → a test file's private fun await (×8); leaflet.js / validate.mjs → unexported Go add / join (55 cross-language rows); Vec::new() → a private fn new in another crate (×7); ui.add(…) (egui) → a private add; f.finish() → HashingWriter::finish in another crate. The Rust breakdown after the trait exemption: 171 instance-method, 8 exact-match calls, 16 references — none onto a trait-impl method.

Not touched: the Go extractor's isExported on methods (a separate extractor defect noted in #1731), the C storage class as a node field, and the local-closure gap behind the JS fail (#1669).

Tests

__tests__/cross-file-visibility.test.ts, whole pipeline over source fixtures, each invisible shape paired with the visible one of identical form: C static in another source file vs non-static vs a header's static inline (kept) vs same-file static with the keyword on its own line; Kotlin private fun apply reached through an SDK-style editor.apply() vs a public fun apply on a typed receiver; Go unexported func across packages vs within the package and an exported func elsewhere; Rust a sibling module's private fn vs a parent's private fn from a child, a trait-impl method through a typed receiver, and a pub fn anywhere. 11 cases. resolution.test.ts, frameworks-integration.test.ts, pr19-improvements.test.ts unchanged: 200 across cross-file-visibility + resolution.test.ts, tsc clean.

🤖 Generated with Claude Code

[danusha2345]

Correction, pushed as 42a9267: the first cut's C rule was wrong for headers, and the betaflight number in the description was mostly that error.

Of the 4,451 rows it removed, 4,306 targeted a static inline in a header — MAVLink's generated mavlink_msg_*.h calling protocol.h's _mav_put_char_array / mav_array_assign_char and each other's _pack / _decode helpers. Those calls are real: a header is textually included, so the function exists in every unit that names it. Only a static defined in a translation unit is local to it.

The rule now applies to candidates in a source file (.c .cc .cpp .cxx .c++ .m .mm) only. Same tree, same base: LOST 145, GAINED 0, header targets in the removed set: 0. The 145 are the shape #1730 describes — STM32 USB class sources onto GD32's usbd_enum.c statics, and the descriptor-table get_device_* case. A test pins the header case (static inline in protocol.h, called from core.c, edge kept). The other four corpora are unaffected by the change (no C).

I'll correct #1730's count there too; the 10% figure in it was counting the header calls.