seed(elixir-pattern-match-refactor): SKLD-bench v2.1 challenge pool (120 challenges)

[ty13r]

SKLD-bench v2.1 challenge pool: elixir-pattern-match-refactor

Seventh and final family being shipped this morning (ecto-schema-changeset #9, ecto-query-writer #10, oban-worker #11, security-linter #12, ecto-sandbox-test #13, phoenix-liveview #14).

Partial-state recovery note

The overnight drafting subagent completed easy/medium/hard tiers + 15 fixtures + 12 goldens before the Max subscription rate limit cut it off before the legendary tier was started at all. This PR completes the family with hand-authored:

• family.json (with known_gaps declaration)
• seed.json — gen 0 SkillGenome with 11 starter variants
• evaluation/score.py — refactor-quality scorer
• evaluation/criteria.json
• evaluation/environment.yml
• challenges/_calibration.json — generated post-hoc

Pool stats

• Total challenges: 120 (rich curve target 150 — 80% of target)
• Tier distribution: easy 35 / medium 47 / hard 38 / legendary 0 ⚠️
• Held-out: 20 balanced across tiers
• Capability coverage: 10 capabilities + 1 foundation = 11 dimensions

⚠️ Known gap: legendary tier empty

The legendary tier has 0 of 30 target challenges. Drafting was cut off before any legendary challenges were written.

One orphaned golden/elixir-pattern-match-refactor-legendary-01.ex file exists (~75 lines of reference output) but no corresponding challenge JSON. Future augmentation can use it as a starting point.

Impact: 120 challenges across easy/medium/hard still provide substantial coverage across 11 dimensions. Champion fitness curves won't have a meaningful "legendary" anchor until augmented.

Per-capability primary coverage

Capability	Primary	Notes
with-expressions ⭐	17	✅
refactor-philosophy (F)	16	✅
defensive-nil-checks-elimination ⭐	14	✅
enum-vs-recursion-choice	13	✅
pipe-operator-flows	12	✅
guard-clauses	10	below 12
recursive-functions	9	below 12
function-head-pattern-matching	8	below 12
binary-pattern-matching-basic	8	below 12
map-and-struct-destructuring	7	below 12
cond-and-if-reduction	6	below 12

⭐ = highest-priority capabilities per the research dossier. with-expressions is the bridge between pattern matching and error handling; defensive-nil-checks-elimination is the most-cited single complaint.

6 capabilities are below the 12-per-cap rich target. All are covered across the remaining 3 tiers. Augmentation is a follow-up.

Score.py validation

Check	Result	Target	Status
Sanity (golden easy-01 defensive-nil elimination)	0.9593	≥0.9	✅
Discrimination (ruby_style_user_service fixture)	0.2703	<0.7 (fail)	✅

Discrimination headroom: 0.69 — excellent.

Score.py approach

This family scores refactor quality via structural counting rather than fixed substring matches:

Positive signals (rewarded):

• Multi-clause function heads (same name, multiple def with different patterns)
• |> pipe operator usage
• with expressions
• when guard clauses
• Map/struct destructure in function heads (%User{id: id})
• List head/tail patterns ([h | t])
• Binary patterns (<<"prefix", rest::binary>>)
• Enum.map / Enum.reduce / Enum.filter

Anti-patterns (penalized):

• if / case / cond keyword counts (total ≤2 for refactored output)
• is_nil() defensive guards
• x && x.field Ruby-style safe-nav pun
• Intermediate temp/tmp/result variables breaking pipe flow
• String.starts_with? / ends_with? instead of binary patterns
• Complex function calls in guards

Research provenance

38 citations across 11 capabilities in research.md. Key sources:

• BoothIQ "150k lines of vibe-coded Elixir" post-mortem: "Claude writes Ruby-style Elixir — if/then/else chains, defensive nil-checking, early returns"
• HN troupo: "writes Java even if it's Elixir"
• HN dnautics: "case functioncall() do nil -> ... end instead of idiomatic if var = functioncall() do"
• Elixir Forum Alex66: "Still correcting if/else chains that should be pattern matches"
• Dashbit, José Valim on idiomatic Elixir

Tier methodology

Heuristic per SEEDING-PLAN.md item 4.

🤖 Generated with Claude Code