feat(spaceship-titanic): add leak-safe bundled evaluator

[Abhishek21g]

Summary

Adds the missing evaluate.py for spaceship-titanic — the only bundled task without a first-party evaluator despite README stating all four ship one.

Builds on the approach in #32 with an additional integrity improvement: results.json uses summary scalars + items[] and never writes held-out Transported labels, aligned with the eval-summary contract in #36.

Why

• Without an evaluator, sia run --task spaceship-titanic cannot feed scores into the self-improvement loop
• Per-item gold in results.json would leak into feedback prompts on current main (Fix: stop leaking held-out reference answers to the meta/feedback agents #36 fixes harness-side; this fixes grader-side)

Changes

• sia/tasks/spaceship-titanic/data/public/evaluate.py — Kaggle-style CSV scoring
• tests/test_spaceship_titanic_evaluator.py — contract + leak guard tests

Test plan

• python -m pytest tests/test_spaceship_titanic_evaluator.py -q — 3 passed
• python -m pytest tests/ -q — 132 passed
• Assert expected absent from written results.json

Related

• Complements feat(spaceship-titanic): add evaluator #32 (same feature area)
• Pairs with Fix: stop leaking held-out reference answers to the meta/feedback agents #36 (leak-safe eval surface)

[Abhishek21g]

Verified locally on Abhishek21g/sia pr/32-spaceship-evaluator: 132 tests passed. Closes the only bundled task missing an evaluator; leak-safe results.json pairs with #36.