Extending the skill graph

The graph in internal/skillgraph/seed.go IS the curriculum and the treasure map: strands are islands, prerequisite edges are the fog of war, skills are dig spots. 56 skills today, grades 3–5, 5 strands.

Adding a skill

Append to the seed slice:

{
    ID:            "kebab-case-stable-id",   // never rename: it's the owner-scoped data key
    Name:          "Kid-Facing Name",        // shows on the map spot
    Description:   "One sentence for the question-generation prompt.",
    Strand:        StrandFractions,          // island; new strands need AllStrands() + StrandDisplayName()
    GradeLevel:    4,                        // 3–5 (grade-2 content lives under G3)
    CommonCoreID:  "4.NF.B.3",
    EstimatedMins: 15,
    Keywords:      []string{"..."},          // steer the LLM prompt
    Prerequisites: []string{"existing-id"},  // at least one, unless replacing THE root
    Tiers:         DefaultTiers(),           // learn: 8 problems @75%, prove: 6 @85%/30s
}

Hard rules

• Exactly one root skill (empty Prerequisites). Validate() panics in init() on cycles, dangling prereq IDs, or duplicate IDs — a bad seed kills every binary and test in the repo.
• IDs are forever: mastery snapshots, events, and spaced-rep state key on skill ID per child. Renaming orphans learner data (the stats/notebook code skips unknown IDs, so old data silently disappears).
• Prereqs may cross strands. Grade level is metadata for map layout and planner ordering — unlocking is purely prerequisite-driven.

Verify

go test ./internal/skillgraph/            # validation + graph tests FIRST
go test ./internal/...                    # planner/session tests exercise the graph

Check the map renders sanely: the new spot appears on its island in the /play UI (grade badge, locked until prereqs mastered) — the saas-e2e skill has the browser flow. No frontend changes are needed for new skills; islands and spots render from the API.

New strands additionally need: Strand constant + AllStrands() + StrandDisplayName() in internal/skillgraph/skill.go (the map and dashboard group by these).

Question generation needs no changes — prompts are built from the skill's Name/Description/Keywords/Tier automatically. Consider whether internal/problemgen/mathcheck.go can recompute answers for the new skill's question style (pure-arithmetic patterns only; non-computable questions pass through silently).