docs(benchmark): record ALCF retry rerun

[JaimeCernuda]

Summary

• update the ALCF demo benchmark report after the transient-provider retry fix
• record the improved result: 14/15 clean passes, 1 expected surfaced error, 0 partial recoveries, 0 failures
• mark the transient-provider follow-up as merged/validated in TASK.md

Evidence

• live backend: http://127.0.0.1:17964
• provider/model: ALCF Metis gpt-oss-120b
• evidence JSONL: tmp/clio-demo-benchmark-alcf-metis-20260524-after-retry.jsonl
• key change: workflow_memory_followup is now a clean pass with six Parquet tool calls instead of a partial recovery

Verification

• uv run python scripts/run_demo_benchmark.py --base-url http://127.0.0.1:17964 --data-dir tmp/clio-benchmark-data --output-jsonl tmp/clio-demo-benchmark-alcf-metis-20260524-after-retry.jsonl --report docs/ALCF_DEMO_BENCHMARK_REPORT.md --case-delay-s 5