[mircealungu]

Roadmap 0b, PR 2 of the series (after #905).

Columns

• study_choice (study, skip, known): written only by learner actions:
  • the "+" button → study
  • "−", or the report dialog → skip
  • "I know this" → known
  • typed-in or audio-lesson words → study
  • OTHER_FEEDBACK in an exercise → skip
  • a learner's "bad translation" → skip, until reports exist (PR 3)
• unfit_reason: why the system keeps a word out. Every place that makes a word unfit now gives one (set_unfit_for_study(session, reason)).
  • Quality heuristics (rare_for_beginner, same_as_translation, too_short, too_many_words, arbitrary_multi_word, bad_contexts) plus separated_mwe: worked out again at every update_fit_for_study. bad_quality_reason returns the first check that fails.
  • validator_gave_up and no_usable_context: last until the next recompute (a new look-up or a star), as today.
  • validator_invalid, replaced_by_fix (Validator fix: take the old word out of study when it survives #896's old word) and duplicate_meaning: stay.

Rule

fit_for_study is worked out from the two columns:

• skip or known keeps a word out.
• study lifts only the quality heuristics. The learner doesn't overrule the system.

fit_for_study stays a column, since the scheduler queries filter on it.

Deploy order

1. Migration 26-10-25--user_word_study_choice.sql. It adds the columns and backfills them. Counts from prod on 25 Oct:

   From	To	Rows
   1	study	6,000 of 6,481. The 481 opt-ins Validator fix: being fit for study is not the learner's opt-in #895's bug forced are left out (roadmap item 0).
   -1	skip	1,458
   -6	known	266
   -100	replaced_by_fix	3
   -100	skip	89
   unfit with a validator "invalid"	validator_invalid	26

2. Deploy the API. The new code reads the new columns.

3. tools/recompute_fit_after_study_choice.py judges the 481 again. It's a dry run by default; --apply writes.

The old user_preference column stays in the table, unused, until a later migration drops it. Anything the old code writes to it between steps 1 and 2 isn't carried over, so keep that gap short.

For review: 29 1 rows have no recorded learner action but aren't validator-fix targets either, e.g. l'enquête, 2021 → 2021, ergonomie → ergonomie. They may come from the edit dialog, which doesn't log the event, or from undo after a report. Following item 0, they keep study.

Tests

New test_study_choice.py covers the rule and which reasons stay and which are judged again. The full core and API suites pass (1299 passed).

🤖 Generated with Claude Code

[github-actions]

ArchLens detected architectural changes in the following views: