fix(rust): resolve use crate::... into real cross-file edges

[BTCB]

The Rust extractor's use_declaration handler was emitting a single edge per use statement to a stem-only node ID like _make_id("types"), which never matched any real node and got garbage-collected by the dangling-edge filter. Net result: zero cross-file edges for Rust workspaces, so every lib.rs / types.rs / model.rs showed up as an orphan weakly-connected component with degree=1.

On a 23-crate real-world Rust workspace, this meant:

• 422 weakly-connected "orphan" nodes (mostly type-definition modules)
• same-file vs cross-file edge ratio of 52:1 (healthy Rust projects should be ~3-5:1)
• edge relation histogram: 8681 calls + 4970 contains + only 4 uses
• orphaned type modules disconnected from their in-crate consumers despite sitting in the same crate — e.g. crates/graduation/types.rs had NO PATH to crates/graduation/manager.rs even though manager.rs contains use crate::types::{ModeTransition, StrategyMode, StrategyRuntime};

Fix

Two-pass resolution, mirroring the Python cross-file resolver's structure:

1. Per-file pass (extract_rust): each use_declaration is parsed and recorded on the result dict's new _rust_uses list. The parser handles:

   • use crate::foo::Bar; (single ident)
   • use crate::foo::{A, B as C}; (brace group + aliases)
   • use crate::foo::*; (glob — skipped,
     can't resolve statically)
   • pub use foo::{A, B}; in lib.rs / mod.rs (relative re-export
     of sibling module)
   • Filters out self::, super::, and external-crate paths.
   • No edge is emitted in this pass — targets live in other files.

2. Cross-file pass (_resolve_cross_file_rust_imports): builds a global label → [(node_id, source_file)] index across all parsed Rust files, then for each recorded use, resolves each imported identifier to a concrete target node. Prefers candidates whose source_file matches the use's module prefix (so use crate::types::Foo picks a Foo defined in a types.rs). Falls back to first non-self candidate if no prefix match.

Produces two new edge relations:

• uses (confidence_score 0.95) for ordinary imports
• reexports (confidence_score 1.0) for pub use in lib.rs / mod.rs

Both are INFERRED, not EXTRACTED, because tree-sitter alone cannot fully verify type identity the way rust-analyzer would — two different crates may define identically-named types and the heuristic picks a best-effort candidate.

Cache schema bump

cache.py now mixes a _CACHE_SCHEMA_TAG = b"v2" constant into every file hash. Pre-existing cache entries (which lack the new _rust_uses field) silently stop matching and get re-extracted on next run. No manual rm -rf graphify-out/cache needed. Bump this tag whenever the extractor output schema changes again.

Specifically on the graduation crate (6 files, shown in commit-msg narrative above): 0 cross-file edges → 15 uses + 17 reexports = 32 new edges. StrategyRuntime, ModeTransition, GraduationScorecard now correctly path to StrategyLifecycleManager in 1-2 hops.

Tests

Added 5 new unit tests in test_multilang.py covering a minimal multi-file Rust fixture (tests/fixtures/rust_crate/):

• test_rust_use_crate_resolves_braced_imports
• test_rust_use_crate_resolves_single_ident_import
• test_rust_pub_use_reexports_in_lib_rs
• test_rust_cross_file_edges_are_inferred
• test_rust_use_crate_never_produces_dangling_imports_from

Full test suite: 417 passed, 0 regressions. (7 pre-existing failures in tests/test_security.py are environment-specific — they reproduce identically on unpatched main when local DNS resolves example.com to a private IP range, unrelated to this PR.)