Record PR42 page-base-flight research evidence

performance/harnesses/glm53_page_base_flight/README.md

@@ -0,0 +1,40 @@
+# GLM-5.3 opaque-page base-flight research evidence
+
+Status: **research-only**.
+
+SparkCache's persistent base-segment I/O mechanism is **implemented** and has
+GPU-free regression coverage. The retained live observations do not qualify
+page-delta restore for GLM-5.3 serving.
+
+## Retained observations
+
+| Condition | Mechanical result | Semantic result | Conclusion |
+|---|---|---|---|
+| Sixteen concurrent 131,072-token page-delta requests with one shared 98,304-token base | Every rank emitted one `sparkcache-page-base-restore-flight/v1` record with 16 participants, one physical base read, and 15 avoided reads | Codeword responses were not reliable | Host-I/O coalescing worked, but the workload exceeded the 20 GiB GLM hybrid-cache residency available per rank and cannot support a correctness claim |
+| Two concurrent 131,072-token page-delta requests within resident capacity | One admitted restored request returned the wrong codeword; a request recomputed without admitted restore returned the correct codeword | The failure occurred with both SparkCache CUDA placement enabled and disabled | The defect is in a path common to reconstructed page-delta restore; the evidence does not isolate CUDA placement |
+| One 131,072-token flat snapshot stored as 13 authenticated macro objects | Restore completed in 1.55–1.70 seconds | The exact `red` codeword was returned | Full-snapshot restore is the verified operational fallback for this evidence scope |
+
+The mechanism evidence proves bounded host reads, authenticated manifests, and
+request accounting. It does not prove that reconstructed page-delta state is
+safe to admit for GLM-5.3 generation.
+
+## Harness contract
+
+`qualification.py` is retained as a deterministic research harness. It uses
+stable single-token codeword oracles, records raw response hashes as diagnostic
+data, and never starts, stops, or restarts a service. A mechanically complete
+verdict has `kind=research-verdict` and `status=research-only`; the harness
+cannot emit a qualification status.
+
+Publication records the rank-0 scheduler-log offset before the base request.
+It submits bounded two-token scheduler steps and reads only later log bytes
+until one `KV Transfer metrics` record reports four ranks and four held
+digests. This readiness check remains required before private-tail requests.
+
+Replay collects every result and the unrelated later request before returning.
+It preserves rejected receipts and exits nonzero after writing them. Manifest
+inspection and bounded rank logs remain required to associate responses with
+the exact page-delta roots and base-flight records.
+
+The 16 × 131,072-token case is retained to reproduce the observed mechanical
+coalescing and residency failure. It is not a supported qualification workload.