feat(tutor): the exercise it plans is the card you get

[mrviduus]

The Tutor calibrates an exercise type for every word from that card's SRS stage — recognition for a new word, recall at stage 2, context for a well-known one with a sentence. It is computed on the server, deliberately not taken from the model's suggestion so the plan cannot schedule a jarring jump, carried to the client with everything needed to render it…

…and then thrown away. buildPlanCard hardcoded options: null and blankSentence: null, so all three types drew the same flashcard and the calibration reached the learner as a coloured badge.

The shape now follows the plan

type	card
recognition	four options, no sentence prompt — which word means this
recall	no options. A flashcard the learner grades against themselves, which is the point of that stage; four options would be a recognition exercise wearing a recall label
context	four options and the saved sentence with the word removed — the cloze, in the sentence the word was actually met in

One implementation of what the choices are

Options come from McOptions, extracted from ReviewCardBuilder — which carried two copies of the same forty lines, one for multiple_choice and one for context. The builder goes from 145 lines to 61, the rule has one home and three callers, and it finally has tests:

• the answer is always among the four, and never also present as its own distractor (case-insensitively — the old pool branch compared case-sensitively and the LLM branch did not);
• the cascade prefers the LLM's distractors → the learner's own vocabulary → a hardcoded list, so a card is never short of options (three choices would themselves tell the learner something);
• malformed distractor JSON costs the card its distractors, never the review.

For the Tutor the pool is the other words in the same session — words the learner has actually met make a real choice rather than obvious filler.

Decisions worth noting

• Clients pick the component from options, not from the type string, so a card can never claim options it does not carry. A correctOptionIndex arriving without options is ignored rather than indexed into undefined.
• reviewMode stays vestigial and unread. Its comment in the shared DTO explicitly says not to build on it, so nothing here does; the comment is updated where it described the old behaviour.
• context with no sentence is pinned even though it is unreachable (CalibrateForStage downgrades it to recall first) — an empty cloze is the one outcome worse than a plain card.
• This is the one place where the card shape is chosen by the server rather than by the client's Blitz/Flashcards preference, and CLAUDE.md now records the exception rather than leaving the two statements to contradict each other.

Verification

• dotnet test tests/TextStack.UnitTests — 1251 pass (6 new McOptions, 5 new exercise-shape)
• web 713, mobile 369 — pass; tsc --noEmit on both — clean; vite build — clean
• dotnet format textstack.sln --verify-no-changes — clean

Not verified in a browser: a Tutor session needs an account with vocabulary and an LLM call, and there is no OpenAI key in the local stack. The projection and the option builder are unit-tested; the two screens are not.

🤖 Generated with Claude Code

https://claude.ai/code/session_011rgEMvYYi4Egj99dVtvm6E