fix: coerce str prompts to Prompt in exemplar selectors

[finitearth]

Summary

• select_exemplars accepted a raw str prompt in practice but passed it straight to task.evaluate, which turned the string into a list of single characters via list(prompts) and later crashed with AttributeError: 'str' object has no attribute 'construct_prompt' (after running predictions on every character first — hence the misleading "fails after many runs").
• Convert str -> Prompt at the selector boundary in both RandomSelector and RandomSearchSelector, mirroring how optimizers coerce initial_prompts in BaseOptimizer.__init__. task.evaluate stays strict (Prompt-only).
• Widened the BaseExemplarSelector.select_exemplars type hint to Union[str, Prompt].

Test plan

• Parametrized the shared selector contract (respects_n_examples, accepts_str_prompt, n_examples_kwarg) over both RandomSelector and RandomSearchSelector
• Added regression test that a raw str prompt is coerced rather than split into characters
• Full suite green: 138 passed, 1 skipped

🤖 Generated with Claude Code

[github-actions]

Tests	Skipped	Failures	Errors	Time
141	0 💤	0 ❌	0 🔥	7.496s ⏱️