fix(add): accept qualified position IDs and make --dry-run resolve placement

[fohara]

Fixes #53.

--before/--after rejected the ID lash prints

lash show and lash list qualify IDs with their file (index#beta-task), so
that is the string people paste back into --before. It failed with "task not
found", which reads as the task being missing rather than the argument being
spelled the way the tool spells it — and --depends-on on the same command
line accepts the qualified form, so one invocation could need both.

Position IDs now go through PlacementResolver::local_position_id, which
strips a file# qualifier when it names the target file and errors when it
names a different one. Accepting the qualifier is not the same as ignoring it:
a qualifier pointing elsewhere means the caller expected the task in another
file, and positioning against whatever local task shares the slug would be
silently wrong.

The qualifier matches the file's @id, its name with or without .md, and any
trailing portion of its path; a #task: prefix on the local part is tolerated
because that is how @depends-on references are written.

The not-found error now names the IDs that do exist at that level, since the
bare "not found" was misleading precisely when the task did exist.

--dry-run did not validate the position

Dry run printed the request back field by field and exited 0. It never opened
the target file, so it passed for a --before naming a task that did not
exist — a false green on the one thing it is for.

create_task now splits into plan_task (load, validate, resolve placement)
and the emit that follows. Dry run calls plan_task, so there is no separate
dry-run path left to drift out of agreement with the real one. It also reports
the resolved insert line rather than the argument it was handed, stated as a
lower bound when the emitter still has to step past a preceding task's
free-text body.

Tests

• 12 unit tests in placement.rs covering every accepted qualifier spelling,
  the cross-file rejection, and the new error content
• 10 integration tests in add_command_test.rs covering both halves end to end,
  including that a failing dry run writes nothing and a passing one writes
  nothing either

Full workspace suite passes; clippy and fmt clean.