test(integration): init results-dir and pass --systemname in real_accumulation tests

[FileSystemGuy]

Summary

`main._main_impl()`'s orgname-resolution gate now requires every gated mode (closed/open/whatif) to have a `mlperf-results.yaml` sentinel pinned into the results-dir by `mlpstorage init`. The universal argparse plumbing additionally requires `--systemname` (or `MLPSTORAGE_SYSTEMNAME`) on emitting commands.

The three module fixtures in `tests/integration/test_real_accumulation.py` predated both, so every subprocess CLI invocation exited 2 with either "`--systemname/-sn` is required" or, for the vectordb/kvcache fixtures, "[E101] results-dir has not been initialized". All 12 tests in the file failed at fixture setup.

Changes

• Define `TEST_ORGNAME` and `TEST_SYSTEMNAME` module-level constants and add a `_canonical_prefix(results_dir, mode)` helper that builds `<results_dir>///results//` — one edit if the layout changes again.
• In each of the three module fixtures: run `mlpstorage init ` before the benchmark invocations, and add `--systemname ` to every benchmark argv.
• Update every path-shape assertion to walk the canonical prefix instead of the (now incorrect) `<results_dir>//...` shape.
• For vector_database, also add the `<index_type>` segment (DISKANN — the default) between engine and command. Current production splits per index_type (Rules.md §2.1.27): on-disk shape is `vector_database//<index_type>///`.
• Production now combines engine + index_type into the metadata's `model` slot (e.g. `milvus_DISKANN`) so the per-index_type workload grouping matches the per-index_type path split. Update the metadata-schema and discovery assertions to expect the combined token.
• The heterogeneous test now symlinks each per-fixture canonical-prefix's per-type subdir into the combined dir, so discovery walks exactly the trees the three fixtures produced.

Test plan

• 6/6 vectordb + kvcache tests pass locally:
  `uv run python -m pytest tests/integration/test_real_accumulation.py -k 'vectordb or kvcache' -m '' -o addopts=` → 6 passed
• The 5 training tests + heterogeneous test (which depends on the training fixture) need ~520GB free disk for unet3d's CAP-01 capacity gate; my dev box has ~133GB. The canonical-prefix change is the same mechanical pattern verified end-to-end on vectordb/kvcache, but a CI machine with sufficient disk should run them.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅