test: validate four-model live GPU PD and numerical controls

opt_prime/examples/pd/http_matrix/README.md

@@ -31,3 +31,17 @@ python examples/pd/serve_batched.py --model "$MODEL" --device cuda:1 --slots 4 \
 ```
 
 The new flags affect the in-process Torch producer only. Nonpositive producer chunks are rejected. Disabling reuse or using smaller chunks can increase prefill work/overhead; these options do not promise a speedup or universal bitwise equality. CPU factory and four-model GPU controls are recorded under issue #241.
+
+### Live GPU PD matrix
+
+The validator accepts `--device`, `--pd-prefill-device` and the producer controls above, records actual producer policy and asserts remote prefill/transfer counters. It also accepts `--http-module` for a trusted local HTTP implementation candidate; the candidate is loaded before its consumers so exception classes retain their identity.
+
+```bash
+git show 9e442cd:opt_prime/opt_prime/accel/serve/llmd.py > /tmp/pd-http-safe.py
+PYTHONPATH=opt_prime python opt_prime/examples/pd/http_matrix/validate.py \
+  --http-module /tmp/pd-http-safe.py --model "$MODEL" --output "$OUTPUT" \
+  --device cuda:1 --pd-prefill-device cuda:0 \
+  --pd-prefill-no-prefix-cache --pd-prefill-chunk 256
+```
+
+Run from the repository root. Omit the producer arguments for native control. `passed` means lifecycle/count assertions passed; concurrent/serial token equality is reported separately, including failed numerical comparisons. See gpu-live-pd240 for the complete default-policy, FP32 and matching-policy results.

opt_prime/examples/pd/http_matrix/STATUS.md

@@ -25,3 +25,7 @@ Structured SDK logs and cancelled/malformed replies previously killed the receiv
 ## Explicit Torch producer policy (#241)
 
 New opt-in controls disable producer prefix reuse and select producer chunk size; defaults remain cache-on/chunk2048. Public cache-off/chunk256 configuration passes all four real GPU PD HTTP paths:48/48 concurrent equality,12/12 native-PD serial equality at128/2048/8192tokens,84remote prefills and0producer hits. Lifecycle checks pass.20focused and464full tests pass (15skipped,135subtests). This is tested numerical reproducibility, not a speedup or arbitrary-input equality guarantee; original default-policy discrepancies remain in#240.
+
+## Live GPU PD matrix (#240)
+
+All four native/PD BF16 lifecycle configurations pass across128/2048/8192tokens. Default PD has35/48concurrent equality versus native48/48, and11/12native-PD serial equality. FP32 restores all divergent models; BF16 explicit cache-off/chunk256 controls#241 restore four-model48/48concurrent and12/12serial equality. Original divergences and validator module-identity failure are retained in evidence240. The self-contained HTTP-candidate loader was fixed and Qwen0.6B rerun passes. No matched speedup or compute/copy overlap claim.

opt_prime/examples/pd/http_matrix/evidence240/events.jsonl

@@ -0,0 +1,3 @@
+{"id": "a4854a4bbae8442ea6c1a5b3ce82e7ba", "ts": "2026-10-23T23:33:18+09:00", "task": "GPU live PD matrix (aicomp #240)", "result": "Qwen3-0.6B BF16 GPU0->GPU1 HTTP lifecycle passes;21remote prefills/44544transferred tokens. Concurrent equality8/12, retained as numerical discrepancy. Remaining native/PD queue49704 active.", "next": "Complete four-model BF16 native/PD matrix, then FP32 controls; preserve all numerical failures.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"dtype": "bfloat16", "decode_slots": "4"}, "artifacts": [{"path": "analysis/serving-matrix-201/gpu240-qwen06b-pd.json", "size": 24146, "sha256": "84de5338ae7a227f"}], "experiments": ["experiments/2026-10-23-gpu-live-pd240.md"], "labels": ["experiment"], "review": false, "todos": [], "note": "", "agent": "codex", "host": "xeon-max", "branch": "main", "commit": "351d39d-dirty"}
+{"id": "bb809b6ecfa54c1c913cdbbaff56c857", "ts": "2026-10-23T23:36:05+09:00", "task": "GPU live PD matrix (aicomp #240)", "result": "Four-model BF16 native+PD8/8 lifecycle runs pass,84remote prefills. Native concurrency48/48 exact; PD35/48 exact(Q06=8,Q4=12,KONI=7,Gemma=8 of12). Native/PD serial11/12; Gemma2K differs.", "next": "Run FP32 and producer-cache-off controls before attributing numerical differences; keep full goal active.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {"dtype": "bfloat16", "decode_slots": "4"}, "artifacts": [{"path": "analysis/serving-matrix-201/gpu240-summary.json", "size": 1555, "sha256": "46799a4baa76b48e"}], "experiments": ["experiments/2026-10-23-gpu-live-pd240.md"], "labels": ["experiment"], "review": false, "todos": [], "note": "", "agent": "codex", "host": "xeon-max", "branch": "main", "commit": "351d39d-dirty"}
+{"id": "75a2497a5045452ca4660b574cf6211f", "ts": "2026-10-23T23:44:56+09:00", "task": "GPU live PD matrix (aicomp #240)", "result": "FP32 controls restore Q06/KONI/Gemma36/36 concurrent and9/9 native-PD serial equality. BF16 cache-off gives36/36 concurrent but7/9 serial; matching chunk256 restores Q06/Gemma6/6 serial. Public-policy4model queue20549 now running.", "next": "Validate public cache/chunk flags, publish evidence and retain originalBF16 mismatches; RNGD long KONI queue remains active.", "state": "active", "done": false, "blocker": "", "metrics": {}, "params": {}, "artifacts": [], "experiments": ["experiments/2026-10-23-gpu-live-pd240.md", "experiments/2026-10-23-pd-producer-cache241.md"], "labels": ["experiment"], "review": false, "todos": [], "note": "", "agent": "codex", "host": "xeon-max", "branch": "main", "commit": "351d39d-dirty"}