[mircealungu]

Builds on #900 (merged).

Problem. validate_and_fix checks the meaning in the word's preferred bookmark, and that is sometimes a generated example. A FIX there moved the learner's word to a different meaning, one that fits the example. Since 1 June, 658 fixes were judged on generated examples. That's 58% of the generated examples the validator judged, against 13% of learners' own sentences, and 395 of those fixes changed the word itself (e.g. mennesker → umanned from "Dronen er mennesker…"). The example was generated for the meaning, so when the two disagree it's the example that's wrong: "selv" = "even" must not become "selv om".

The same goes for a sentence the validator calls BROKEN_SENTENCE (#900). It says nothing about the word, which may fit any other sentence.

Change. In both cases (any non-VALID verdict in a generated example, or BROKEN_SENTENCE in any sentence), _try_another_sentence:

• leaves the learner's word on its meaning;
• if the sentence was a generated example, records it in the new example_sentence_misfit table;
• judges the meaning again in another sentence: another sentence of the learner's own (a reader fragment too: it is the only place a wrong meaning can be corrected, and the fragment swap keeps it out of exercises), else the next unused example. When the learner's own sentence was the broken one, examples only. At most 2 sentences in all (worst case 6 validator calls per word).
• With no sentence left, gives the word up. A meaning that only ever failed as FIX in examples is logged as [word_not_fit]; broken sentences as [broken_sentence]. The meaning itself is not touched.

Examples recorded as misfits are never offered again: the fragment swap (replace_fragment_with_example) and /alternative_sentences both skip them.

A FIX in the learner's own sentence still changes the meaning, as before, because that's the sentence they actually looked the word up in.

Deploy: run tools/migrations/26-10-25--add_example_sentence_misfit.sql before deploying. It creates a new table, so no existing table gets rebuilt, but the example queries read from it.

Tests: 6 new DB tests:

• a FIX in an example keeps the meaning and drops the example;
• with no sentence left, the word is given up on without touching the meaning;
• a VALID in an example is kept;
• a broken own sentence goes to an example, not to another own sentence, and the word stays in study;
• a broken own sentence with no example gives up without touching the meaning;
• after a bad example, a reader fragment of the learner's own is judged.

Full suite: 1261 passed, 1 failed. The failure is test_cloze_alternatives::test_a_translation_verdict_is_not_a_cloze_verdict, which passes on its own (3/3) and next to the new tests, and doesn't touch this code. It looks order- or fixture-dependent.

Not in this PR:

• The 395 learners' words already moved to a different meaning by generated examples need reviewing and possibly restoring. That's a prod data job; I'll show the query and the batch before any write.
• Exercises can still show other generated example bookmarks via the arrows, and those are never checked. Checking every sentence shown is the larger design.

🤖 Generated with Claude Code

[github-actions]

ArchLens detected architectural changes in the following views: