fix(rust): resolve a forward-referenced type to its local declaration - #3903

What does this PR do?

Fixes #3782.

A Rust struct, enum or trait used above its declaration got a sourceless stub: ensure_named_node only resolves names already in seen_ids. _rewire_unique_stub_nodes folds it back only for a corpus-unique label, so a same-named type in another file leaves the edge on the stub.

extract_rust now pre-scans the file's struct_item/enum_item/trait_item names before walk() (the Verilog, SQL and Bash pre-scan idiom), and ensure_named_node resolves against them.

• Only these three items are pre-registered: a type alias or union without an impl never becomes a node, so its references would be dropped as dangling.
• The scan never goes deeper than walk() (no function, static or const bodies), so every registered id becomes a node.
• Option 2 (the stub's origin_file in the rewire) would cover other extractors too; a natural extract.py follow-up.

Type of change

• Bug fix
• New feature
• Documentation
• Tests or CI
• Refactor
• Security fix

Verification & Invariants

Invariant: a reference above a same-file struct, enum or trait, spelled the same and alone on its casefolded id, resolves like one below it; anything else keeps v8's per-file result.

• Read the CONTRIBUTING.md guide.
• Reproduced the issue and identified the invariant.
• Made the smallest fix necessary.
• Added a regression test (if bug fix) or isolated boundary test.
• Kept the PR description synchronized with the final implementation.
• Documented any limitations / unsupported cases explicitly.

Casefold collisions fail closed: when fn handle, const HANDLE, struct HANDLE or impl Tr for HANDLE shares struct Handle's id, or the reference spells HANDLE, it keeps the stub v8's extractor gives it; references after the id's owner are unchanged.

Known limitation: resolution is flat per file and keeps only the last path segment, inheriting its errors: qualified paths (io::Error, imp::Child), Self::X without alias, generic parameters, mod scopes. On tokio 2 of 85 moved references are wrong (imp::Child); on 351 crates of a local cargo registry, roughly 85 of 2,279 (201 in the suspect classes). Pre-scanning only unqualified names loses 57+ correct ones. Failing closed withholds 29 likely correct links (e.g. type declared before its colliding function).

Side effect in the final graph: some targets reached a real node on v8 only via a forward-reference stub sharing their id, later folded onto the type; they now stay unresolved. imports_from: 53 wrong mio use …::io bindings to a private Io, 12 tokio ones to Scheduler, 5 correct ones (e.g. socket2 pub use sys::SockFilter). References: 8 wrong jobserver HANDLE bindings to struct Handle, the only crate that drops (77 → 69); +2,271 = 2,279 moved − 8.

How was this tested?

# issue repro (`graphify update .`)
v8 d6eaa8a:   src_engine_a_before -> sink                (sourceless stub)
this branch:  src_engine_a_before -> src_engine_a_sink

# 7 new tests (tests/test_multilang.py)
v8 d6eaa8a:   4 failed, 3 passed (the 3 negative-space tests pass on both)
this branch:  7 passed

# typed edges landing on a real node, per crate, on v8 then on this branch
python -c "import sys,tempfile,pathlib as p;from graphify.extract import extract;d=p.Path(sys.argv[1]);r=extract(sorted(d.rglob('*.rs')),cache_root=p.Path(tempfile.mkdtemp()),parallel=False);s={n['id'] for n in r['nodes'] if n.get('source_file')};print(sum(e['relation'] in('references','implements','inherits') and e['target'] in s for e in r['edges']))" <crate>
tokio 1.52.3: 2433 -> 2518; mio 1.2.0: +1; socket2 0.6.3: +0
same command over 351 crates of a local cargo registry: +2,271

pytest tests/         -> 6104 passed, 15 skipped (v8: 6097 passed, 15 skipped)
ruff check .          -> All checks passed
pyright <both files>  -> 0 errors (v8: 0)

Removing any of these checks (bodies, modules, collisions, impl, exact name) turns a new test red.

Graphify-specific checklist

• I updated generated skill artifacts (uv run python -m tools.skillgen --bless) when changing their source fragments.
• I confirmed that AST/structural extraction remains deterministic (no ambient state dependencies like ENV variables).
• I reviewed changes for security implications (no unsafe interpolation into shell/Python).
• I confirmed no API keys or local-only graph data are included.
• (If applicable) I disclosed AI authorship in my commit messages.