test: add real P/D reference trace and topology feature A/B result

benchmarks/traces/README.md

@@ -47,13 +47,14 @@ Optional fields (zero-filled when absent):
 
 ## When to capture a new trace
 
-Three reference traces are included, covering different workload patterns:
+Four reference traces are included, covering different workload patterns:
 
 | Trace | Prompt length | Contention | Samples | Covers |
 |-------|--------------|------------|---------|--------|
 | `sharegpt-h200` | short (median 165) | moderate (median 12) | 3,400 | Real conversations, moderate load |
 | `chatbot-synthetic-h200` | long (median 6,497) | extreme (median 93) | 3,060 | Long prompts, stress testing |
 | `bimodal-h200` | mixed (median 3,411) | extreme (median 116) | 3,400 | Mixed short+long, stress testing |
+| `kermit-pd-topology-h200` | mixed (median 1,942) | high (median 11) | 1,500 | **Only trace with real P/D disaggregation** — `pod_type` and topology field (`topology_distance`) populated; the other three are monolithic-only |
 
 Capture a new trace when testing features that depend on:
 

benchmarks/traces/kermit-pd-topology-h200-spec.yaml

@@ -0,0 +1,46 @@
+workload:
+  name: kermit-pd-topology-h200
+  trace_type: real
+  data_source: ShareGPT + long synthetic + shared-prefix (combined)
+  model: Qwen/Qwen3-32B
+  hardware:
+    gpu: NVIDIA H200 (141GB)
+    gpus_per_replica: 1
+    platform: CoreWeave CKS (Kermit), Kubernetes
+    topology: single-region (US-EAST-01), uniform InfiniBand fabric
+  serving:
+    stack: llm-d-router EPP (topology-extractor, topology-affinity-filter/scorer,
+      predicted-latency-producer) + vLLM P/D disaggregated (NixlConnector, kv_both)
+    epp_plugins: topology-extractor, topology-affinity-filter, topology-affinity-scorer,
+      predicted-latency-producer, prefix-cache-scorer, queue-scorer
+    max_model_len: 8192
+    prefix_caching: enabled
+  load:
+    generator: inference-perf
+    profile: ShareGPT + long-synthetic (2k-8k tokens) + shared-prefix
+    api: completion (streaming)
+    stages:
+      - {rate_qps: 1, duration_s: 300}
+      - {rate_qps: 2, duration_s: 300}
+      - {rate_qps: 4, duration_s: 600}
+  topology_finding:
+    # Evaluated topology_distance as a raw XGBoost feature via
+    # benchmarks/offline_feature_ab.py (ordinal hop-distance encoding). Result:
+    # no distinguishable signal — the improvement is within seed noise and the
+    # tool's self-test cannot separate it from overfitting gain. Only zone and
+    # region tiers appear here and they transfer at near-identical speed, so
+    # distance carries little latency signal on this fabric. No A/B artifact is
+    # shipped; the consuming approach is per-class residual correction, not a raw
+    # feature (see llm-d-latency-predictor PR #3 and predictor issue #30).
+    verdict: neutral
+    seeds: 10
+    split: 90/10 train/test per seed
+  known_limitations:
+    - single-region, uniform InfiniBand fabric — only zone- and region-level
+      distance observed (no host, rack, or cross-region pairs)
+    - topology_distance "host" stratum shows 0 samples despite the router's
+      topology-affinity-scorer always selecting a same-host prefill candidate when
+      one was available — unresolved observability discrepancy, see llm-d-router PR
+    - this file is a stratified downsample (1,500 of 16,186 real entries) to stay
+      under the repo's 1MB trace-file limit; full trace available on request
+    - prefix_cache_score is near-zero for 99.7% of samples