Select interview problems adaptively and maintain a rolling study plan

[rosshd]

Parent: #56

Depends on: the interview-profile, interview-skill-graph, problem-catalog, and attempt-lifecycle children of #56.

Problem

The current curated drill selector checks whether a problem tag appears in the topic's current focus or goal and otherwise returns the first problem.

That is not enough to produce an efficient interview study plan.

The next activity must balance prerequisites, weak patterns, delayed review, hint dependence, difficulty, target level, available time, and novel transfer.

Goal

Add a deterministic adaptive problem selector and interview study planner that can explain every recommendation.

Selector inputs

• Interview profile, target, time horizon, language, and session budget.
• Skill graph prerequisites and learner evidence.
• SRS due state and delayed-retrieval failures.
• Attempt outcomes, time spent, hint exposure, solution exposure, and abandonment.
• Problem metadata, revision, difficulty, expected duration, supported language, and skill coverage.
• Recent problem history and near-duplicate relationships.
• Learner requests such as "practice graphs", "mock interview", or "review weak areas".

Selection policy

• Filter invalid, unsupported, unavailable, or rights-restricted candidates first.
• Respect blocking prerequisites without over-drilling one concept indefinitely.
• Prefer due retrieval and unresolved weaknesses when the learner requested normal study.
• Interleave patterns after initial acquisition.
• Use novel transfer problems after assisted or familiar success.
• Avoid immediate repeats and near-duplicates unless repetition is intentional.
• Keep challenge near the selected mastery profile and adjust from observed evidence.
• Fit the expected activity into the available session.
• Reserve model judgment for ambiguous semantic ranking, not basic eligibility or safety.
• Return a concise reason and alternatives with every recommendation.

Study plan

• Produce a rolling plan rather than a brittle fixed syllabus.
• Show current focus, next few activity goals, due reviews, and readiness gaps.
• Recompute after material new evidence while preserving completed history.
• Distinguish target-date urgency from open-ended durable learning.
• Represent uncertainty rather than inventing readiness precision.

Acceptance criteria

• The same input state produces the same eligible candidate set and baseline ranking.
• Every selected problem has a learner-visible explanation grounded in current evidence.
• Unsupported languages, missing prerequisites, invalid sources, and unavailable content are excluded.
• Assisted success schedules novel transfer instead of granting immediate broad mastery.
• Repeated failure triggers bounded remediation or prerequisite work under Define repeated-miss progression without false mastery #48 rather than endless same-problem retries.
• The selector avoids accidental near-duplicate transfer checks.
• Session-length constraints change the recommendation appropriately.
• Learner overrides are honored and recorded without corrupting the underlying readiness model.
• Fixtures cover new, advanced, time-constrained, hint-dependent, stale, and target-date learners.
• Property or invariant tests cover eligibility, prerequisite, and repeat-avoidance rules.
• make check passes.

Out of scope

• Claims that a plan predicts a specific employer's exact questions.
• Global collaborative filtering.
• Opaque model-only ranking.
• Gamified streak pressure that overrides learning needs.