feat(eval): add quality provider adapter (#156)

[luongnv89]

Closes #156

Summary

PR 2 of 5 from the Skillgrade integration plan. Wraps the existing static SKILL.md linter (src/evaluator.ts) in the EvalProvider contract introduced by PR 1 (#155) — without modifying src/evaluator.ts. If this adapter needed ugly workarounds, the contract would be what had to change; it didn't, so the interface from PR 1 holds.

Approach

Thin delegation. qualityProviderV1.run() calls evaluateSkill() and maps EvaluationReport onto EvalResult. applicable() is a cheap stat(skillMdPath) — no other gating, since the evaluator surfaces every other issue as a finding.

Mapping (EvaluationReport -> EvalResult)

EvaluationReport	EvalResult
overallScore	score
grade !== "F"	passed
categories	categories (1:1 on id, name, score, max)
topSuggestions	findings with severity: "info"
original report	raw (stable per schemaVersion)

Evaluator's free-form findings: string[] and suggestions: string[] stay in raw — the adapter intentionally does not invent a string→Finding conversion the issue didn't ask for.

Changes

File	Change
src/eval/providers/quality/v1/index.ts	New — EvalProvider implementation
src/eval/providers/quality/v1/fixtures/well-formed.json	New — snapshot for pass-path corpus skill (score 91, grade A)
src/eval/providers/quality/v1/fixtures/missing-frontmatter.json	New — snapshot for fail-path corpus skill (score 21, grade F)
src/eval/providers/quality/v1/index.test.ts	New — 12 tests (contract, applicable(), snapshot per corpus skill, registry integration, mapping invariants)
src/eval/providers/index.ts	Registers quality@1.0.0 via registerBuiltins()
src/eval/providers/index.test.ts	Updated to expect 1 registered provider (was 0 in PR 1)
tests/fixtures/skills/well-formed/SKILL.md	New corpus fixture
tests/fixtures/skills/missing-frontmatter/SKILL.md	New corpus fixture

src/evaluator.ts unchanged — git diff HEAD~1 src/evaluator.ts returns zero lines.

Non-determinism handling

Snapshot tests strip these before deep-equal against checked-in JSON:

• startedAt, durationMs — stamped by the runner on every call
• raw.evaluatedAt — wall-clock set by evaluateSkill()
• raw.skillPath, raw.skillMdPath — absolute paths depend on checkout; fixtures store a __FIXTURE__/<name> marker

Drift in fixtures is a review artifact — if evaluator scoring changes, snapshots flag it so reviewers can decide whether the change is intentional.

Test Results

• bun test src/eval/ — 80 pass, 0 fail (67 existing PR 1 tests + 12 new quality provider tests + 1 existing registerBuiltins test)
• bun run typecheck — clean
• Full bun test src/ — 1413 pass, 5 fail (pre-existing unrelated failures in publisher.test.ts and cli.test.ts, called out in the issue; not addressed here)

Acceptance Criteria

• quality provider registered and resolvable via registry.resolve("quality", "^1.0.0")
• Snapshot test per corpus skill passes — adapter output matches checked-in JSON
• src/evaluator.ts is untouched (diff confirms zero changes)
• bun test passes (all new tests green; pre-existing unrelated failures as called out in issue)

Boundary kept for PR 3

src/cli.ts does not import providers/index.ts or call registerBuiltins() — PR 3 owns that wiring. No user-visible behavior change in this PR.