fix(extract): parse .vue SFC <script> with the right grammar

[papinto]

Summary

.vue Single File Components were extracting almost nothing because the whole file was being parsed with the wrong grammar.

.vue is dispatched to extract_js, which picks a tree-sitter grammar by file suffix. Since .vue is neither .ts nor .tsx, it fell through to the JavaScript grammar, and the entire SFC — <template> markup, <script>, and <style> — was handed to that grammar. Markup isn't valid JS, so the parse produced a top-level ERROR node and no imports, symbols, or type references were recovered.

This adds a dedicated extract_vue that parses only the <script> block, with the grammar its lang implies.

Impact

Measured on a 300-file random sample from a production Vue 3 + Vite + TypeScript app:

Metric	Before	After
Files extracting nothing (0 edges)	40%	6%
Files with no import edges	69%	8%
Total edges	2,834	8,882 (3.1×)

The residual ~6% are genuinely template-only components (no <script>), which correctly produce just a file node.

How it works

1. Mask, don't strip. Every character outside a <script> body — template, style, and the <script …>/</script> tags themselves — is replaced with a space, while newlines are preserved. The surviving script therefore sits at its original line numbers, and the blanked regions parse as empty. No line-offset bookkeeping required.
2. Pick the grammar from lang. tsx → TSX, js/jsx → JS, and ts or unspecified → TS. TypeScript is a strict superset of JavaScript, so defaulting an unannotated block to the TS grammar is lossless and matches the dominant Vue 3 + <script setup> convention.
3. Recover dynamic imports. A small regex pass picks up import('...') calls (e.g. defineAsyncComponent(() => import('./X.vue'))) that the AST pass does not emit as edges — the same rescue extract_svelte/extract_astro already do.

This goes a step further than the Svelte/Astro extractors, which regex-scrape import strings only: because Vue's component logic lives in <script>, parsing it with a real grammar recovers the full symbol and type-reference graph (classes, functions, typed props referencing imported types), not just imports.

.vue files also now join the JS cross-file symbol-resolution pass. They were previously excluded from it precisely because the whole-file parse failed; with a working per-file parse, an SFC's calls resolve to definitions in other files like any .ts file's would.

Implementation notes

• _extract_generic gains an opt-in keyword-only source_override: bytes | None = None. When provided, it parses those bytes instead of reading the file, while still keying every node/edge off the real path. Default-off, so no existing caller changes behavior.
• The one-site and path.suffix != ".vue" guard in _collect_js_symbol_resolution_facts is removed. .vue stays in the JS-family suffix set used elsewhere for cache-bypass.

Tests

tests/test_vue_extraction.py (11 cases): masking + line-number preservation, <script setup lang="ts"> static imports, symbol extraction with correct lines, typed props referencing an imported type, dual <script> + <script setup> blocks, dynamic-import recovery, plain-JS blocks, template-only files, and end-to-end cross-file call resolution.

The JS/TS import-resolution, Svelte/Astro, and symbol-resolution suites remain green.

[MentorFilou]

The same issue was similarly addressed in #1141 . Sadly the author didnt respond for 3 weeks.

While yours seems to have better test coverage, @safishamsi suggested that reusing the svelte function is the right call, even asking to not create a distinct vue wrapper function! Consider reading their response on that PR to evaluate possibly changing this here.

Hoping to see this merged soon 👍

[safishamsi]

Thanks @papinto — great catch and the right fix. Feeding the whole SFC to the JS grammar produced a top-level ERROR and dropped everything; masking the non-<script> regions with spaces (preserving newlines so line numbers stay accurate) and parsing the script with the lang-named grammar is exactly the correct approach, and wiring .vue into cross-file symbol resolution is a nice bonus.

Landed on v8 in 349465b with your authorship preserved. One maintainer fix on top: the <script> open-tag scan now skips over quoted attribute values, so a > inside one — Vue 3.3+ generic components like generic="T extends Record<string, unknown>" — no longer ends the tag early and swallows the body. Added a regression test for it. Verified end-to-end: a <script setup lang="ts"> SFC now recovers its imports and symbols with no ERROR node. Closing as merged-by-port — thanks for the contribution!