fixtures: a corpus with ground truth, and the defect it found

[inquerium]

Adds test/fixtures/, which docs/OPENCLAW.md already requires and which did not exist. Eight synthetic learners, each built to exhibit one property of the mastery model, each carrying an expect block stating what is true of it by construction.

Independent of #7. Touches nothing under world/.

Why

Every claim about the model was checkable only by reading the code. Nothing that examines the record, the scheduler, or a generated activity could be measured against anything.

test/fixtures.test.ts holds the corpus to its own claims. A persona that quietly stops exhibiting the property it is named for does not fail loudly. It makes everything measured against it report success against nothing.

The personas

id	what it is by construction
cold-start	Eight years old, empty record. Placement opens.
cold-start-young	Five years old, empty record. Placement is withheld on purpose.
stuck-blending	Nine attempts, two correct. Sorts to the top of the queue.
lapsed-summer	Three crammed skills and one well-spaced one, same child, same summer off.
crammer	Twenty right answers in one sitting.
vowel-confuser	Reads every short e as short i, with short i mastered. Deliberately unlabelled.
accommodated	Three active accommodations.
above-band	Six years old doing grade-2 place value.

Two properties are now arithmetic rather than assumption. Eight correct answers across widening gaps buy a half-life near 65 days, so a well-spaced skill survives a summer and only lapses around 150. The same eight crammed into six consecutive days lapse in 60. lapsed-summer carries both in one child, so the spacing effect is visible as a difference rather than asserted.

vowel-confuser carries a real misconception in the response column with no misconception row. A detector scored against a record that already names the answer has found nothing.

The defect it found

nextTargets tested opportunities >= 6 && p < 0.5 before it tested lapsed. Six opportunities is an ordinary number of tries to learn something, so a child who mastered a skill in spring and took the summer off came back reported as stuck: the tutor was told to change modality or drop to a prerequisite for a child who needed a two-minute review of something she used to know.

peak_p_known is the only thing that separates "lost it" from "never had it". statusOf already reads it. The stuck branch now consults it.

stuck-blending still reports stuck, which is the case the branch is for.

Checks

npm test passes, 193 tests. CI cannot run: the upstream account is billing-locked and every workflow is refused before it starts.

🤖 Generated with Claude Code