feat(vba): record each procedure's error policy (#287)

* feat(vba): record each procedure's error policy

Nothing in the graph recorded how a procedure handles errors, and neither
of the two questions that matters is greppable: an ABSENCE scoped to a
procedure body has no text to match, and whether an `On Error Resume Next`
is ever closed depends on the whole body rather than on any one line. In
this corpus that hides 816 procedures with no protection at all and 516
procedures whose suppression scope runs to the `End Sub`.

Every procedure's `function` node now carries an `errorPolicy` object:
`protection`, `handlerLabel`, `handlerStartLine`, `handlerEndLine`,
`behavior`, `handlerCount`, `resumeNextOpen` and `danglingTarget`.

Decisions taken:

- ZERO new node kinds and ZERO new edge kinds. Section 4 of
  `docs/vba-error-handling-plan.md` rejects a `label` node plus a
  `handles-error` edge because 96.5% of the corpus's 3,911 line labels are
  the same label (`errores`) doing the same job — one bit, which belongs in
  a field. Measured on the corpus: 14,649 nodes / 21,361 edges / 26,613
  unresolved references before and after, byte-identical.

- A label defined but never targeted by an `On Error GoTo` is CONTROL FLOW,
  not a handler. 137 labels in the corpus are exactly that (`siguiente`,
  `salir`, `fin`). Reporting one as a handler would give a confidently wrong
  answer to "does this procedure handle errors", the worst failure mode
  available here, so a label only opens a handler region when an `On Error
  GoTo` in the SAME procedure names it.

- Every rule scans the MASKED line, so `s = "On Error GoTo errores"` sets
  nothing (the #209 discipline), and the procedure body's END is the
  existing `PROCEDURE_END_RE` — the same constant the `proc-end` rule in
  `enums-consts.ts` dispatches — rather than a second end detector, which is
  what makes the colon-separated single-line procedure form (#208) work here
  for free. The body's START is the node the procedures pre-walk already
  emitted, so there is no second `PROC_RE` dispatch either.

- The classifier agrees with `scripts/vba-coverage-probe.mjs` by
  construction: same label regex, same statement-keyword guard, and the same
  treatment of `On Error GoTo 0` / `-1` as resets and of every other target
  as a handler label. Its `protection` census reproduces the probe's exactly
  — 3,774 handler / 227 resume-next / 816 none over 4,817 procedures.

- `On Error GoTo -1` is valid VBA (it clears the current error) and appears
  0 times in the corpus. It is treated as a reset and pinned by a fixture so
  the zero-occurrence case cannot rot untested.

- `behavior` is deliberately left null. It is derived in E3, where the
  handler body's calls are already being classified; guessing it here would
  be fabrication.

The classifier's `count` stays at 0 for its whole life, so a file whose only
content is an `On Error` statement is still a file with no symbols and gains
no module node.

Closes #259

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019gmKKUq1ng5ESk6Qhxu77d

* docs(vba): correct the E2 dangling-target acceptance to the measured 0

The E2 acceptance line predicted exactly one dangling handler target in the
corpus, carried over from the pre-probe hand census. E1 landed the committed
probe and its reconciliation table already records the measured figure as 0;
this line was the last place still asserting 1, so the implementer of E2 was
being asked to find something that is not there.

The corpus does contain a handler literally named `noExiste`, defined in its
own procedure and therefore correctly not dangling — almost certainly what the
hand census mistook for an unresolved target.

The protection distribution on the same line is updated to the measured
3,774 / 227 / 816 for the same reason.

Refs #259

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019gmKKUq1ng5ESk6Qhxu77d

---------

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

Author: ardelperal

CHANGELOG.md

@@ -21,6 +21,7 @@ and adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 - Every parameter a VBA procedure declares is now its own symbol in the graph, linked to the procedure that owns it and to the class it is typed as, so you can ask which procedures take a given class and follow it from there. (#257)
 - SQL that a form or control is bound to at runtime — assigned to `RowSource`, `RecordSource`, `ControlSource`, `Filter` or `OrderBy` — now shows the tables it reads, whether the statement is written inline or built up in a variable first. (#252)
 - Access domain functions such as `DLookup` and `DCount` now record the table or saved query they read, so a procedure that touches data only through them no longer looks like it touches no data at all. (#255)
+- Procedures now record how they handle errors — whether they have a handler, silently suppress errors, or have no protection at all — so unguarded code paths can be found without reading every module. (#259)
 
 ### Changed