docs: verify issue #600 visual scenes dataset

Add a verification note and regression test showing that
training/data/prompt-enhancement/visual-scenes-500.jsonl already
exists on main with 500 valid records.

docs/issue-600-verification.md

@@ -0,0 +1,41 @@
+# Issue #600 Verification
+
+Status: already implemented on `main`.
+
+Issue: Prompt Enhancement: Visual Scenes — 500 Terse→Rich Pairs
+
+What is already present on `main`
+- `training/data/prompt-enhancement/visual-scenes-500.jsonl`
+- 500 JSONL records
+- every record includes `terse`, `rich`, and `domain`
+- every `domain` value is `visual scenes`
+- 500/500 full records are unique
+
+Evidence gathered from a fresh clone
+- `python3` validation against `training/data/prompt-enhancement/visual-scenes-500.jsonl` returned:
+  - `count = 500`
+  - `unique_records = 500`
+  - `unique_terse = 435`
+  - `domains = ['visual scenes']`
+  - `missing_keys = 0`
+- remote branch `fix/600` still exists from closed PR #731 (`feat: 500 visual scene prompt enhancement pairs (#600)`)
+- SHA-256 of `training/data/prompt-enhancement/visual-scenes-500.jsonl` on `origin/main` matches the same file on `fix/600`, which shows the requested dataset is already present on `main`
+
+Verification commands
+```bash
+python3 - <<'PY'
+import json
+from pathlib import Path
+path = Path('training/data/prompt-enhancement/visual-scenes-500.jsonl')
+records = [json.loads(line) for line in path.read_text().splitlines() if line.strip()]
+print('count', len(records))
+print('unique_records', len({json.dumps(r, sort_keys=True) for r in records}))
+print('unique_terse', len({r['terse'] for r in records}))
+print('domains', sorted({r.get('domain') for r in records}))
+print('missing_keys', sum(any(k not in r or not str(r[k]).strip() for k in ('terse', 'rich', 'domain')) for r in records))
+PY
+```
+
+Recommendation
+- Close issue #600 as already implemented on `main`.
+- This branch only adds a durable verification note and regression test so the zombie issue can be closed cleanly without regenerating duplicate training data.