Markdown importer silently drops all requirements after a fenced code block that directly follows a heading

[koppor]

Bug description

When a fenced code block's opening fence directly follows a Markdown heading (with only blank lines in between), the importer misparses the fences and silently ignores every specification item after the block's closing fence. There is no error or warning — trace reports ok - 0 total and exits 0, and impl->req coverage tags pointing at the dropped items show up as orphaned.

Minimal reproduction

spec.md:

# Spec

## A diagram before any requirement

```mermaid
graph TD
  A --> B
```

## Some requirement

`req~example~1`

A requirement that should be imported.

Needs: impl

$ java -jar openfasttrace-4.9.0.jar convert -o specobject spec.md
<?xml version="1.0" encoding="UTF-8"?>
<specdocument></specdocument>

$ java -jar openfasttrace-4.9.0.jar trace spec.md
ok - 0 total

Expected: req~example~1 is imported (1 item). Actual: no items, no diagnostic.

The info string is irrelevant — ```mermaid, ```json, ```foobar, and a bare ``` all trigger it. What matters is the position: move the same block so a paragraph of text precedes it and the requirement imports fine.

Root cause

In MarkdownImporter.configureTransitions(), the TITLE state (entered after a heading) has no transition for MdPattern.CODE_BEGIN. The opening fence therefore matches TITLE → START via EVERYTHING, and the closing ``` then matches CODE_BEGIN from START (the \w* info string is optional), opening a phantom CODE_BLOCK. Since LineParserStateMachine keeps its state on unmatched lines, everything up to EOF (or the next stray fence) is swallowed.

This looks like a gap in the fix for #280 (#433), which added the CODE_BLOCK state but only wired CODE_BEGIN from START.

Versions

Reproduced with openfasttrace 4.9.0 (latest release), OpenJDK 21, Linux. Originally surfaced in JabRef via the org.itsallcode.openfasttrace Gradle plugin 3.2.0.

Impact

Silent requirement loss: spec authors get zero signal that items vanished from tracing. JabRef currently works around it by moving diagrams after the last requirement.

I'll open a PR adding the missing TITLE → CODE_BLOCK transition plus a regression test.