No cheap first-pass spellcheck on generated text: 'dyrpark' shipped to A1 learners

[mircealungu]

Problem

We have no cheap first-pass check on generated text, so the two things we do run
are all-or-nothing: either an expensive LLM pass, or nothing at all.

Concrete evidence, found while preparing the EuroCALL talk. A production A1
simplification of a dr.dk article (article 4894028) opens:

"Sommeren var meget varm i Europa. En dyrpark i Frankrig havde ti
krokodilleæg."

dyrpark is not a Danish word — it should be dyrepark. It shipped, in the
version served to A1 learners, who are precisely the readers least able to spot
it. A plain spellchecker catches that in under a millisecond.

Why this is the "Escalate to the LLM" pattern

The pattern is in our own catalogue
(https://llm-patterns.mircealungu.com/) and we have not applied it here:
run something cheap over everything, and spend the expensive check only on the
residue.

Today the grammar-correction pass is the only instrument, and it is an LLM call
per article. That is why it was added, disabled, and re-argued about
(7cafb9b7 Disable Haiku grammar correction pass after simplification) - the
cost is per-article regardless of whether anything is wrong.

Compare zeeguu/core/language/language_check.py, which gets this right for a
different question: lingua is nearly free, answers a closed question ("is this
plausibly Danish?"), and only the ambiguous cases need more. Spelling should
work the same way.

Do we have a multilingual spellchecker?

No. Nothing in requirements.txt and no spellcheck code in zeeguu/.

Coverage is not a problem. We offer 11 languages - da, nl, fr, de, it, sv,
pt, es, ro, en, el - and every one of them has a mature Hunspell dictionary.
(The language table holds 31 rows, but most are not live; do not size this
against that list.)

Hunspell via spylls (pure Python, no build step) or cyhunspell is the
obvious choice. pyspellchecker bundles only ~7 languages and would leave gaps;
symspellpy is faster but needs a frequency list built per language, which is
work we do not need to do given the dictionaries already exist.

Design notes

• The goal is triage, not correction. We are not trying to fix the text; we
  are deciding whether it is worth an LLM call. So false positives are cheap
  (one wasted call) and false negatives ship a typo. Tune accordingly.
• Proper nouns will dominate the noise. News text is full of Bjørnøya,
  place names, brands. A raw "any unknown token" trigger fires on everything, so
  the signal has to be a share of unknown tokens over a threshold, and probably
  ignore capitalised tokens mid-sentence.
• Where to run it: after simplification, before the article is stored -
  the same place the grammar pass sits.
• Worth measuring first: run hunspell over a sample of stored simplifications
  per language and see what the unknown-token rate actually looks like before
  picking a threshold.

Not scheduled

Filed for later; no work planned right now.