Add regression tests for the sentence-correction tap-and-retype flow

Investigating a manual bug report ("typing petits as the correction
for petit doesn't validate"): both a direct test of
sentence_correction.tscn and a full exercise_host integration test
(mission_02_pont_des_phrases) tap the faulty token, type its exact
expected correction, and submit -- both pass, confirming the
validator and the scene wiring correctly accept "petits" when the
word is tapped once and the correction is typed before pressing
Valider.

Caught along the way: a first draft of the direct-scene test used a
lambda connected to answer_submitted to capture its parameters into
outer local variables -- GDScript closures capture locals by value
at creation time, so reassigning them inside the lambda never
reaches the outer scope. Switched to GUT's
assert_signal_emitted_with_parameters, the supported way to inspect
signal arguments in a test.

No fix applied yet: these tests confirm the intended interaction
works, but don't rule out a real fragility noticed while reading the
code -- toggling a token off deletes its LineEdit (and whatever was
typed in it) immediately, so an accidental second tap on the same
word silently discards the correction. Needs the user's exact click
sequence to confirm before changing behavior.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

tests/exercises/test_sentence_correction_scene.gd.uid

@@ -0,0 +1 @@
+uid://c7si2246mty63