Create eval suite with test fixtures and okf:test skill

[siculo]

Problem

No fixtures, runner, or CI gate exist. Skills are actively iterated but regressions are discovered only during live sessions.

Proposed Change

Directory structure: evals/

evals/
├── README.md
├── okf-create/
│   ├── 01-single-source/
│   │   ├── input.md                   ← sample source document
│   │   ├── expected-concepts.yaml     ← expected list: path, type, title
│   │   └── expected-log-entry.yaml    ← expected log.md entry
│   └── 02-multi-source/
│       ├── input-a.md
│       ├── input-b.md
│       └── expected-concepts.yaml
├── okf-update/
│   ├── 01-new-concept/
│   │   ├── bundle-snapshot/
│   │   ├── input.md
│   │   └── expected-changes.yaml      ← NEW: 1, UPDATES: 0, SPLITS: 0, CONFLICTS: 0
│   └── 02-split/
│       ├── bundle-snapshot/
│       ├── input.md
│       └── expected-changes.yaml
├── okf-reconcile/
│   ├── 01-stale-detection/
│   │   ├── bundle-snapshot/
│   │   ├── sources/
│   │   └── expected-report.yaml
│   └── 02-new-in-sources/
│       ├── bundle-snapshot/
│       ├── sources/
│       └── expected-report.yaml
├── okf-validate/
│   ├── 01-missing-type/
│   │   ├── bundle-snapshot/
│   │   └── expected-errors.yaml       ← ERRORS: 1, WARNINGS: 0
│   └── 02-conformant/
│       ├── bundle-snapshot/
│       └── expected-errors.yaml       ← ERRORS: 0, WARNINGS: 0
└── okf-search/
    ├── 01-full-text/
    │   ├── bundle-snapshot/
    │   └── expected-results.yaml      ← rank 1: path, title
    └── 02-type-filter/
        ├── bundle-snapshot/
        └── expected-results.yaml

New skill: skills/test/SKILL.md

A new okf:test skill that:

1. Parses optional arguments skill:<name> and evals:<path>.
2. Discovers all fixtures under the evals directory.
3. Runs each fixture against the skill's classification/proposal logic on a temporary copy — never writes to the real bundle.
4. Reports pass/fail with full diffs for failures.

Example output:

OKF Skill Eval
─────────────────────────────────────────
Fixtures run : 10
Passed       : 9
Failed       : 1

FAILED
  ✗ okf-update / 02-split
      expected: SPLITS: 1
      actual:   SPLITS: 0
      note: the skill classified the concept as UPDATE instead of SPLIT

On failure, the skill prints the full diff and does not auto-fix anything, since failures indicate a regression in the skill logic that requires manual review.

Files Affected

• New directory: evals/ with fixture structure above
• New file: skills/test/SKILL.md

Estimated Effort

~4–6 hours to create initial fixtures; ~1 hour for the test skill.