AnswerValidator.validate_sentence_correction compared the typed
correction to the expected word with raw string equality, with none
of the tolerance already built for dictations: no case folding, no
typographic-apostrophe or whitespace normalization, no decomposed-
accent handling. Tapping "petit" and typing "petits" could fail for
reasons having nothing to do with the grammar point being tested
(e.g. a stray capital from autocapitalization), which violates the
"an ambiguous answer is never auto-marked wrong" invariant.
Found this while chasing a manual bug report: an automated test
reproducing the user's exact click sequence (tap once, type, submit)
did NOT reproduce the failure, which actually pointed at the real
gap -- the logic worked for byte-identical input, but had zero
tolerance for anything else. Now reuses TextNormalizer (built for
dictation analysis) before comparing, case-insensitively, while still
showing the child's raw typed text in the feedback. Three new tests
cover case, incidental whitespace, and decomposed accents.
Rebuilt both exports so the shipped binaries include the fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
scripts/correction/: AnswerValidator (multiple choice, sentence
correction via tap-the-faulty-word + retype) and DictationAnalyzer
build the CorrectionResult structure from cahier des charges §6.
Dictation analysis aligns words with WordAligner (Wagner-Fischer edit
distance) before any character comparison, so an omitted or added
word doesn't cascade into false errors on every following word.
TextNormalizer handles non-breaking spaces, typographic apostrophes,
and decomposed-vs-composed French accented characters via an
explicit table (GDScript has no built-in NFC/NFD normalization).
Error classification follows the data annotations only (rule_id
present -> grammatical, absent -> unclassified), never a runtime
linguistic guess.
scripts/pedagogy/: DifficultyManager (per-skill difficulty from
recent results), AttemptFlow (3-attempt graduated aid: hint, then
explanation, then reveal, matching exercise-lifecycle.md's second
gap), and ReviewScheduler (the five mastery states from §9, with
session-counted spaced review intervals [1, 3, 6, 12] rather than
wall-clock dates, since a family device's clock isn't trustworthy).
All of it is pure: no scene tree, no clock or RNG read directly,
every effectful input (session number, attempt outcome) is a
parameter. 65 headless tests across both layers, including the
word-omission-doesn't-cascade regression and the full
decouverte -> acquise -> a_reviser -> acquise path.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Denis Defolie