feat(eval): corpus #19 — 3 v0.2 tracked-refs tasks (DRAFT spec)

[w1ne]

Summary

First slice of workstream #19 (corpus expansion) per the v0.2-to-v1.0 gap-closure roadmap. Adds 3 v0.2 tracked-refs corpus tasks under eval/tasks/, each exercising a natural-form case that requires v0.2 capability (canonical face refs surviving transforms or unambiguous booleans):

• fillet-translated-box — box(...).translate(...).fillet(r, { face: 'top' })
• subtract-then-fillet-rim — box(...).subtract(cyl).fillet(r, { face: 'top' }) (the canonical "subtract.fillet" case from the gap-closure roadmap)
• chamfer-rotated-wedge — box(...).rotate([1,0,0], θ).chamfer(d, { face: 'top' })

Each task ships with prompt.md, solution-expert.kcad.ts, and harness.ts. New eval/corpus-v0.2.test.ts mirrors the seed-task sanity pattern (eval/runner.test.ts) and asserts each expert solution scores 100% via runTask + MockAgentClient.

Spec + plan are marked DRAFT v1 in their bodies (docs/superpowers/specs/2026-05-03-corpus-v0.2-tracked-refs-design.md, docs/superpowers/plans/2026-05-03-corpus-v0.2-tracked-refs.md). The brainstorm step was skipped — the gap-closure roadmap text was specific enough that the spec is a focused operationalization rather than open exploration. Open questions for controller review are listed in the spec's "Open questions" section (3 vs 5 tasks, difficulty band assignments, tilt-correctness rubric tightness, naming).

Why this slice

The dispatch order doc (PR #56) calls for Wave 1 = #19 + #21 + #22 in parallel. This is the first concrete #19 deliverable: the v0.2 module's 3-task slice (range "3–5" per the gap-closure roadmap §Corpus design). Three tasks rather than five keeps the slice reviewable; the missing 0–2 are punted to a follow-up pending controller feedback on the rubric shape.

Test Plan

• npm run typecheck — clean
• npm run build:cli — clean
• npm test — 903 pass, 25 skipped, 0 failed (was 870 before; +3 corpus tests + a few transitive expansions)
• Each expert solution evaluates clean via kernelcad evaluate --json
• Each expert solution scores gate_pass=true, score=1 via runTask+MockAgentClient
• Branch rebased on develop tip (3a309be)

Scope notes

Included: 3 task folders + 1 vitest sanity check + CHANGELOG [Unreleased] entry. package-lock.json updated to reflect the v0.2.0 version bump (the lockfile was stale at 0.1.0 in the worktree until npm install ran).

Out of scope:

• Real-API agent runs (corpus-author-time concern; lives at G1 pre-flight per the roadmap).
• Difficulty-band wiring beyond per-task self-classification (no harness behavior depends on bands yet).
• Tasks for v0.3+ modules (those modules haven't shipped).

Companion PR

Depends on PR #56 landing (the gap-closure scaffolding) for the corpus-design spec it references. If #56 is held for review, this PR can stand alone — it doesn't import any of #56's files at runtime.