Function build_prompt 

pub fn build_prompt(
    lesson: &Lesson,
    submission: &Submission,
    results: &ExecResults,
) -> Prompt

Build the LLM feedback prompt for a graded submission.

Pure (§2.1): it reads only the borrowed domain values and performs no IO, env read, or network call, so identical inputs always render byte-identically. The layout is a fixed structure — the task, the lesson’s success_criteria when present and non-blank (ADR-0020), a single fenced copy of the submission, a captured-output section, and a check-results section (one line per outcome, labeled with its lesson.checks entry by index) — not the lesson’s llm_evaluation_prompt template (ADR-0006). results.outcomes is expected 1:1 with lesson.checks (the shape run_checks produces); the rendering never silently drops a verdict if they desync (see render_checks). Every interpolated value is neutralized, so the fences and labels each appear exactly once even when the submission forges them: injected text can never be read as code or as a verdict.