[fohara]

Replaces #39, which GitHub closed when its base branch was deleted on merge of
#40. Same change, rebased onto main.

What was wrong

Appending after a task carrying a hand-written @depends-on: a, b, c put the
new task two lines too far down, outside the ## Tasks section entirely. In
the repro the ## Notes heading ended up separated from its own prose.

count_annotation_lines added one line per parsed dependency. That holds for
what MarkdownEmitter::format_task_annotations writes, one reference per line,
but the parser splits a comma-separated value into several references off a
single line, so the count ran two lines long.

When the task is last in the file the emitter's clamp hides this, which is why
it only bites when content follows.

@labels had the same problem from the other direction and was simply skipped,
with a comment saying it could not be counted, because labels can be inline on
the checkbox line or in an @labels: block and the parsed metadata does not
record which.

The fix

Both are one defect: the number of source lines cannot be recovered from parsed
metadata, because the same metadata can be written in more than one shape. So
record it at parse time. The parser already collects a task's annotation lines
and therefore already knows the count; it now keeps it on
Task::annotation_line_count, and the resolver uses that.

This settles the comma form, @labels, and any future annotation with more
than one written shape, without the resolver having to know about any of them.

Tasks built in memory rather than parsed (the TUI's in-memory tasks, test
fixtures) have no source lines to have counted, so the field is 0 and the
resolver falls back to deriving a count from the metadata on the emitter's
one-line-per-annotation assumption. That keeps existing behaviour for callers
that never touch a file.

Tests

Three integration tests appending after a task with comma-form @depends-on,
with one @depends-on per line, and with block-form @labels, each asserting
the new task lands inside ## Tasks and the following section keeps its
content. Two unit tests covering the recorded count winning over the derived
estimate, and the fallback for an unparsed task.

The comma-form and block-labels tests were confirmed to fail without the fix.

Also filed here

Two lash format bugs found while verifying this, filed in lash.index.md
rather than fixed here:

• lash format deletes every section after ## Tasks. A file with ## Notes
  and ## References comes back with only the header and the tasks. Exit code
  0, no warning.
• lash format duplicates inline labels on every run, so #docs #infra
  becomes #docs #infra #docs #infra and grows without bound. format --check
  reports the file as needing formatting forever.