feat(vba): emit variable nodes for module-level state

[ardelperal]

What changed

Module-level VBA state is now in the graph, both as symbols and as relationships.

Nodes. Every module-level Public / Private / Global / Dim / Static declaration — including WithEvents fields — emits one variable node with qualifiedName: '<Module>.<name>', a visibility folded from the declaration keyword, metadata: { declaredType, isArray, isWithEvents, isConst: false }, and a contains edge from the module/class node via the existing pending-source mechanism. Procedure locals emit nothing.

References. Inside a procedure body, an identifier that matches one of that same file's module-level variable names emits an UnresolvedReference tagged synthesizedBy: 'vba-module-var', property-set for a write and property-get for a read. The direction comes from the control sweep's existing isDirectAssignment predicate, which is exported rather than duplicated; a leading Set is stripped from the prefix before the call so Set gblConn = New Foo reads as the write it is.

Resolution. These references bind only to a variable node in their own file and never fall through to global name matching — m_count and strSQL are declared privately in dozens of modules, and a global match would invent couplings between modules that never reference each other.

Decisions worth reviewing

• Visibility does not reuse foldVisibility. That helper folds unrecognised keywords to public, which would claim every Dim gblFoo is exported. In a declarations section Dim and Static mean module-private; only Public and Global export.
• References de-duplicate per (procedure, variable, direction). A procedure that reads a global on twenty lines couples to it once. Reads and writes stay separate rows, so the direction question is still answerable. First occurrence keeps the line number — the same trade-off the Me.<Control> sweep already makes.
• Shadowing uses a dedicated per-procedure name registry, not localVarTypeMap. The type map's bare-Dim path yields to an outer declaration of the same name, and it never sees parameters at all, so both cases would have read as unshadowed. Parameters are parsed off the signature line, which also closes the shadow hole for Sub Guardar(ByVal codigo As String).
• A declarations-only .bas now gets a module node. Module-level variables count toward hasAnySymbols, exactly as module-level Const already did.

Tested

__tests__/extraction-vba-module-variables.test.ts — 26 tests, one per acceptance criterion that is expressible as a unit test, plus the two regression pins:

• Public gblConn As DAO.Database → one variable node + contains, and the DAO references edge still emitted exactly once
• Dim x As Long inside a Sub → no node, no reference
• module variable + procedure local of the same name → the local produces no node, and a read inside that procedure produces no module-var reference, while a procedure without the local still reports its read (fix(vba): localVarTypeMap has no procedure scope — qualified calls resolve to the wrong class #205 holds); covered for a typed local, a bare untyped local, and a parameter
• Public WithEvents m_Form As Form_X → node with isWithEvents: true, subscribes-event edge unchanged
• x = gblConn → property-get; gblConn = Nothing and Set gblConn = New Foo → property-set
• visibility folding for all five declaration keywords, isArray, multi-variable lines, re-declaration, string-literal names, other.gblConn member access, and file-scoped resolution

npx tsc --noEmit is clean. npx vitest run: 3261 passed, 21 failed, 59 skipped.

Not verified

• The 21 failures are pre-existing and environmental, not from this change. 19 are the documented Windows EPERM temp-directory teardown quirk (worktree-detection, multi-repo-workspace, and two extraction.test.ts submodule cases — all fail in afterEach rmSync, with their assertions passing). The other 2 are npm-sdk cases that resolve a real installed platform bundle on this machine instead of their fixture. No VBA extraction or resolution suite fails. I did not run a main baseline to prove this, because the shared stash and sibling worktrees made a clean checkout unsafe here — the evidence is the failure sites, which are all in files this change does not touch.
• The corpus-wide acceptance numbers are not reproducible in this repo. nodesByKind['variable'] ≈ 2,900, the drop in proceduresWithNoOutgoing, and the 15-reference precision spot-check all need the Access corpus, which is not checked in here. Someone with the corpus should run the T0 probe before this is called done. The mechanism the proceduresWithNoOutgoing criterion depends on is in place: the probe counts a procedure as a leaf only when it has zero edges and zero unresolved references, and a procedure that only touches globals now has references.

Closes #251

🤖 Generated with Claude Code

https://claude.ai/code/session_019gmKKUq1ng5ESk6Qhxu77d