test(dataset): offline unit tests for sample_minibatch and LCB parser (#88)

[luciferlive112116]

Summary

• Add offline unit tests for sample_minibatch (without/with replacement, validation, determinism) and load_tasks determinism.
• Add tests for _parse_lcb_tests, including JSON-string payloads and the stdin/stdout alias pair.
• Normalize stdout as an output alias in _parse_lcb_tests, mirroring existing stdin handling.

Fixes #88

Context

Continues the offline unit-test contribution line:

• test(envfile): add secrets env loader unit tests #19 — envfile loader tests
• test(roles): add postprocess truncation unit tests #15 — postprocess truncation tests
• test(roles): offline unit tests for prompt assembly (SPEC §4.4) #38 — prompt assembly tests (SPEC §4.4)
• test(orchestration): offline unit tests for gather_in_batches #41 — gather_in_batches orchestration tests
• fix(benchmarks): livecodebench load() delegates to load_tasks correctly (#86) #87 — livecodebench facade fix

No overlap with open dataset PRs (#45 MMLU split, #66 loud toy fallback) or open test PRs (#73–#75, #52–#57).

Test plan

• pytest tests/test_dataset.py -v (9 passed)
• Pure stdlib / offline — no network or API keys required

[tmimmanuel]

The parser tweak in src/trinity/orchestration/dataset.py:330 is consistent with the existing LiveCodeBench test shape, and the new tests cover the intended behavior for sample_minibatch() and _parse_lcb_tests().

Residual risk: test_load_tasks_is_deterministic_and_truncates() still exercises load_tasks() directly, so it is only fully offline if the Hugging Face fetch path falls back as expected. That is acceptable here, but it leaves a small environment-dependent test gap.