Adds a minimum-resource basic tier to the disagg perf-sanity configs: one ctx
server, one gen server, one round, one request, for each (GPU, model) pair the
directory covers. Ten configs and ten test-list entries.

Why

Every config under tests/scripts/perf/disaggregated pins a server count and a
parallel shape chosen for throughput. The cheapest still takes two nodes; the
DeepSeek-V4-Pro ones take six to nine because they run 5-8 workers. There is no
tier that only answers "does this model still start, transfer KV over NIXL, and
serve a request" -- so that question currently costs a full perf run.

How the parallel shape was chosen

Not by arithmetic. Computing it from weight bytes over per-GPU memory yields a
shape nobody has run, and being wrong costs an OOM on eight GPUs an hour into a
SLURM job.

Each config instead inherits the worker_config of an existing config in this
directory verbatim -- the one whose ctx and gen GPU counts are both the
smallest for that (GPU, model). Every parallelism, KV, MoE backend and
cache-transceiver setting therefore already runs in CI today.

That a single config holds both minima is a property of the current directory, not
an assumption: it holds for all ten groups, and the generator raises rather than
splicing a ctx from one file with a gen from another if a future config breaks it.

Only four fields differ from the base:

num_prompts is concurrency * multi_round in ClientConfig, so each serves
exactly one request. output_length is deliberately left alone: this tier is
about node count, not wall clock, and a short output length starves the
per-iteration statistics the gen worker reports.

What it costs

29 nodes total. Only DeepSeek-V4-Pro shrinks against its own base (5 gen servers
-> 1); the rest were already one server per role and are cheap because they take
the smallest shape rather than the fastest one.

Checks

• submit.py's own get_hardware_config() parses all ten and reports
  num_ctx_servers=1, num_gen_servers=1 for each.
• Every generated file diffs against its base in exactly the four fields above --
  worker_config is byte-identical.
• check-yaml, validate-test-lists, test lists format and waive list check
  pre-commit hooks pass on the new files.

Naming

basic sits between the GPU and the model, matching where this directory already
puts scenario markers (gb200_wideep_..., gb200_stress-...). Putting it after
the model would split a model's configs apart in a sorted listing.

wideep and stress variants are excluded rather than shrunk: a one-request
minimum-node version of either is not the same test, and every model they cover
already has a plain config in the same group.

The generator that produced these, and its 36 tests, live in the CI repo rather
than here.

Dev Engineer Review

• Added ten basic disaggregated configurations for GB200 and GB300 systems.
• Each configuration uses one context server, one generation server, one round, and one request.
• Existing worker settings remain unchanged.
• YAML, hardware parsing, generated differences, test lists, and waiver lists passed validation.
• The scope excludes wideep and stress variants.

QA Engineer Review

• Added ten QA entries in tests/integration/test_lists/qa/llm_perf_multinode.txt.
• Added coverage for all new GB200 and GB300 configurations.
• No test functions were added, modified, or removed.
• CBTS coverage data is unavailable.

Verdict: needs follow-up