fix(vba): classify unambiguous statement-form calls as calls

[ardelperal]

Closes #265

What changed

An unresolved statement-form Sub call was always pushed as unqualified-ident, the bucket for ambiguous bare-identifier reads. That classification is correct for exactly one of the three statement shapes:

Form	Can it be a Const read?	referenceKind
Escribir	Yes — a bare identifier read	unqualified-ident (unchanged)
Call Escribir	No — the Call keyword is only valid on a procedure	calls
Escribir 1, 2	No — a constant cannot take an argument list	calls

detectStatementCall now returns { name, unambiguous }, and the unresolved-reference push site picks calls or unqualified-ident from that flag. Nothing else about the row changes — same fromNodeId, same referenceName, same vba-statement-call-unresolved stamp — so the two kinds can only trade rows with each other.

Resolved calls are untouched: they emit a calls edge and never reach this branch.

One extra piece the issue did not call out

maskStringContent blanks string literals to spaces before the statement detectors see the line, which reduces Escribir "texto" to the identifier plus trailing whitespace — indistinguishable from a bare read. Implementing "an argument list is present" on the space-masked line would therefore have left the single most common statement-call shape in the wrong bucket.

The statement/qualified/With detectors now run on a second, index-identical mask that fills literals with _ instead of spaces, so an argument list stays visible while the literal's contents stay unparseable. Column-sensitive scanners (scanCallSites, scanMeControlReferences, TempVars, DoCmd, Forms-bang) keep the original space mask; maskStringContent's default is unchanged.

Measurement

Two production Access projects (00_EXPEDIENTES, 00_GESTION_RIESGOS), 246 .bas / .cls modules. Baseline is a detached checkout of main at a0345ae.

Extractor output — the committed probe

npm run probe:vba -- "C:/00repos/codigo/00_EXPEDIENTES/src" "C:/00repos/codigo/00_GESTION_RIESGOS/src"

referenceKind	Before (main a0345ae)	After	Δ
calls	7,679	8,976	+1,297
unqualified-ident	1,459	162	−1,297
sum of the two	9,138	9,138	0
references	2,057	2,057	0
member-with	1,488	1,488	0
qualified-call	399	399	0
property-get	215	215	0
property-set	76	76	0

diff between the two probe reports is 8 lines — those two rows and nothing else. Declared procedures 3,840, stub nodes 1,597, every node kind, every edge kind and synthesizer, the top-30 stub targets, SQL tables (69), forms (132) and errors are all identical.

The before column reproduces the figures the orchestrator quoted for current main exactly.

After the resolver — failed vs declined-runtime

These are the numbers the noise-ratio table in docs/vba-reference-kinds.md publishes. Same two projects, indexed separately through the real pipeline.

reference_kind	Total (before → after)	failed (before → after)	failed / total (after)
calls	6,194 → 6,351	693 → 777	~12%
unqualified-ident	166 → 9	85 → 1	~11%
combined	6,360 → 6,360	778 → 778	~12%

Both sums are invariant — rows were reclassified, none created or dropped. Every other kind is identical (member-with 1,401, references 826, qualified-call 380, property-get 215, property-set 76).

Docs

docs/vba-reference-kinds.md:

• the calls row keeps Call LimpiaBuffer and now also names the argument-list form;
• the unqualified-ident row states explicitly that a bare identifier with no Call keyword and no arguments stays there because of the Const-read ambiguity, with a short table and the reasons (FR-3.1 const-first disambiguation, and the name-only DAO-enum / intrinsic-constant gates from enh(resolver): DAO enum values + VBA intrinsic constants should be classified as built-ins #188 that key on this kind);
• the v1.13.0 noise-ratio rows for the two affected kinds are marked superseded and a re-measured section is added, citing both the probe and the indexed run. The 00_VBA_TOOLKIT_BENCH corpus that produced the v1.13.0 numbers is no longer present on disk (its index is empty and it holds no sources), so the re-measurement is on the two Access projects above and says so;
• stale line-number cross-references to src/types.ts and the two extractor files corrected.

Regression guards

• FR-3.1 (fix(extractor-vba): classify unresolved_refs.reference_kind by syntactic shape (round-3) #108) — a bare identifier that names a module Const stays unqualified-ident, including when unambiguous calls sit on neighbouring lines. The existing FR-3.1 test in extraction-vba-reference-kind.test.ts is untouched and still passes.
• fix(resolver): VBA stdlib allowlist misses statement-form calls (MsgBox statement form) #192/fix(resolver): classify VBA statement-form stdlib calls as runtime #195 — MsgBox "x", Shell "calc" and friends still land in declined-runtime, verified through a real CodeGraph.indexAll() round-trip. They move to the calls kind, and the stdlib gate already accepts calls by name, so their status is unchanged. Bare DoEvents stays unqualified-ident.
• enh(resolver): DAO enum values + VBA intrinsic constants should be classified as built-ins #188 — the DAO-enum and intrinsic-constant gates are name-only on unqualified-ident; those are all bare reads and none of them move. extraction-vba-runtime-constants.test.ts passes unchanged.
• A genuine missing callee written Call Foo now comes out calls / failed — the actual point of the issue, since it previously sat behind the const/intrinsic gates meant for bare reads.

One existing expectation was updated, not weakened: extraction-vba-stub-resolver.test.ts Test 9 pinned MsgBox "hi" and Shell "calc.exe" to unqualified-ident. Their kind is now calls; the assertion this test exists for — all three are declined-runtime — is unchanged.

Verification (on main a0345ae merged in)

• npx tsc --noEmit — clean
• pnpm run build — clean
• pnpm exec vitest run vba extraction-sql-query sql-query-discovery — 48 files passed, 747 tests (47 on main; the new test file makes 48, nothing else moved)

Merging main produced one conflict, in CHANGELOG.md, where two ### Fixes bullets landed on the same line — both kept. src/extraction/vba/calls.ts (#245) and src/extraction/vba/text-utils.ts (#247) auto-merged cleanly; both were re-read after the merge and the measurements above were re-run on the merged tree.

New tests

__tests__/extraction-vba-statement-call-kind.test.ts — 13 tests, real files, no mocking: the detectStatementCall flag itself (including the _-mask case and the trailing-whitespace non-case), the three shapes end-to-end through VbaExtractor, the resolved Call Escribir edge, the FR-3.1 guards, and the declined-runtime guards through a real index.