[liuzhengyang699]

random_sample_num_edges samples with replacement until the graph reaches the requested edge count. If the candidates are exhausted or the remaining edges would create cycles, it keeps retrying forever. For example, the two candidates a -> b and b -> a cannot produce a two-edge DAG. This can hang the random-graph baseline in experiments/crosswords/evaluate.py.

Visit candidates once in random order and stop when the target is reached. If this pass cannot reach the target, raise a ValueError describing the sampled and requested edge counts. Existing edges still count toward the target, and the input graph is left unchanged.

Validation: python -m pytest test/swarm/graph/test_edge_sampling.py -q — 5 passed, no LLM calls. Three regressions fail before the fix; the tests bound sampler calls so the old implementation cannot hang the test run. git diff --check passed.