Fixes the ground truth for the `colours_present` task.
**Problem.** The generator built each 20-entry yes/no vector by iterating a Python `set`, so `data.json` listed the answers in the set's hash order (`blue, fuchsia, green, ivory, …`) instead of the order given in `rules.txt` (`black, gray, brown, maroon, …`). 183 of 200 records were wrong, and every model scored 0–7.5% regardless of its perception.
**Changes**
- `colours_present/data.json`: labels reordered to match `rules.txt`. Images are unchanged. Each corrected label was checked against the colours actually present in its image (200/200).
- `colours_present/eval.py`: gold labels are now read from `data.json` (matched on `id`) instead of from the answer file. Existing answer files, which carry the old labels, are therefore scored correctly without re-running inference, and the script prints a note when it sees them. A new optional `--gold` flag points at another `data.json`; output format is unchanged. Verified on 58 existing …
