Adds units 6-39 and regenerates 0-5, so the whole course now uses what the
platform's parser actually supports: \, a stable ID: per activity,
REFERENCE from comprehension exercises to their reading passage, inline gap
answers, and a real INPUT mode per exercise instead of 'either' (which the
parser silently turned into speak-only).
Each conversational dialogue is followed by a role-play replay: the learner
performs one character's lines from meaning while the script is blurred, and
the role alternates between Lara and her interlocutor down the unit.
The two revision units could not be converted in one pass — the model stops
producing after roughly a dozen activities, and the tail of the unit went
missing without any error. They are built from three-lesson chunks and
stitched, which is why 18 and 39 carry 65 and 80 activities.
1367 activities across 40 modules: the production parser accepts all of them,
the linter is silent, and the fidelity checker reports no missing vocabulary
row, dialogue line, exercise item or answer key.
Co-authored-by: Cursor <cursoragent@cursor.com>
