Research bounded flat-page object prefetch

README.md

@@ -356,6 +356,19 @@ mapped arena before submitting its copy spans. A flat 813,068,464-byte
 snapshot therefore requires 13 payload objects rather than 512 logical-chunk
 files; the manifest remains the atomic visibility point.
 
+Bounded flat-object prefetch is **research-only**. The implementation
+authenticates up to four version 2 objects concurrently in request-private host
+buffers and then copies them into mapped placement arenas in manifest order.
+Its GPU-free integrity, ordering, concurrency, and memory-bound tests pass.
+
+The exact GLM-5.3 TP4/DCP1 serving evaluation for SparkCache
+`eabe7fd0c878db7384ef87fe80a1e96b9bedcf67` structurally verified all four
+rank-local 131,072-token snapshots but returned `spark` instead of the expected
+`red`. An equivalent recomputation returned `red`. Consequently, the
+four-reader implementation is not a deployable restore path and does not
+replace the single-reader qualification. See the
+[immutable research receipt](evidence/glm53-flash-dflash7-bf16/flat-v2-four-reader-semantic-rejection-eabe7fd.json).
+
 Flat macro publication and its SparkCache CUDA restore path are
 **implemented and GPU-free tested, not live qualified**. The object-count
 geometry above follows the format contract; it is not a claim of measured

deploy/deepseek_v4/tp4_profile.json

@@ -4,7 +4,7 @@
   "cache_model_profile": "deepseek-v4-fp8-hma",
   "published_runtime_base": "ghcr.io/fujitsupolycom/gb10-vllm-serving@sha256:6fc26fdad81a18f0fff67ce0a05f6d90165625ea2e1cac8a6f39bfb462017028",
   "sparkcache": {
-    "source_sha256": "6651f2823c816fac93779cbca54a8f19c0ed262830953149f3a87d189d1f833b"
+    "source_sha256": "15d62a07088d1212bfea60cecae868844cb582a3ac5273189a24532617ac5590"
   },
   "model": {
     "repository": "deepseek-ai/DeepSeek-V4-Flash-0731",

deploy/glm52_35bpw/profile.json

@@ -5,7 +5,7 @@
   "published_runtime_base": "ghcr.io/fujitsupolycom/gb10-vllm-serving@sha256:6fc26fdad81a18f0fff67ce0a05f6d90165625ea2e1cac8a6f39bfb462017028",
   "base_image_requirement": "exact GLM-5.2 3.5-bpw R7 image recorded by the source container inspection",
   "sparkcache": {
-    "source_sha256": "6651f2823c816fac93779cbca54a8f19c0ed262830953149f3a87d189d1f833b"
+    "source_sha256": "15d62a07088d1212bfea60cecae868844cb582a3ac5273189a24532617ac5590"
   },
   "model": {
     "repository": "brandonmusic/GLM-5.2-EXL3-TR3v4-3.5bpw-MTP78",

evidence/glm53-flash-dflash7-bf16/flat-v2-four-reader-semantic-rejection-eabe7fd.json

@@ -0,0 +1,76 @@
+{
+  "schema": "sparkcache-glm53-flat-v2-reader-research/v1",
+  "status": "research-only",
+  "conclusion": "semantic-rejection",
+  "runtime": {
+    "image_id": "sha256:df4e09a32cdbf1c0e69cc7c4c9e95d890d6c7a1e3eaac84f969912a16fd27dd3",
+    "sparkcache_commit": "eabe7fd0c878db7384ef87fe80a1e96b9bedcf67",
+    "sparkcache_tree": "d88a65ea265a6f212367baa8c4a4970079d6b08a",
+    "sparkcache_source_sha256": "d9a7800ce201b0671676fc8d71423947c7b24e4797758db727c06e0e684495fe",
+    "topology": "TP4/DCP1",
+    "publication_schema": "snapshot-v1",
+    "flat_manifest_schema": "sparkcache-page-snapshot-manifest/v2",
+    "maximum_parallel_reads": 4
+  },
+  "stored_context": {
+    "context_digest": "b4161571df103395e2abae10372a90f35468561ec6c42bf4a7b7f0d0dfda5873",
+    "prompt_sha256": "965acd85cb28f804ab59cdc160688b04efaee14341e0bd27b647673e652ab812",
+    "tokens": 131072,
+    "encoded_bytes_per_rank": 813068464,
+    "objects_per_rank": 13
+  },
+  "structural_restore": {
+    "ranks": [0, 1, 2, 3],
+    "all_ranks_verified": true,
+    "read_and_hash_ms": {
+      "minimum": 484.1,
+      "maximum": 528.8
+    },
+    "placement_ms": {
+      "minimum": 323.7,
+      "maximum": 330.9
+    },
+    "arena_wait_ms": {
+      "minimum": 293.2,
+      "maximum": 297.7
+    },
+    "final_completion_ms": {
+      "minimum": 129.5,
+      "maximum": 131.6
+    },
+    "cache_service_ms": {
+      "minimum": 1231.7,
+      "maximum": 1331.2
+    }
+  },
+  "semantic_restore": {
+    "expected": "red",
+    "observed": "spark",
+    "passed": false,
+    "response_sha256": "0bb1366f58973e94b3cd518d4981be67c86ee35768039e392bb5c494b27bc58e"
+  },
+  "recomputation_control": {
+    "relationship": "one-token-changed prompt with the same token count",
+    "prompt_sha256": "4bb683a895caaaacb783294e65cb9c4b59c808c1e7b563a48edb6cd52b302dfe",
+    "tokens": 131072,
+    "expected": "red",
+    "observed": "red",
+    "passed": true,
+    "elapsed_seconds": 55.14106,
+    "response_sha256": "2c68d02422a6c4bdb42bd10221940894e746342bef6a56695fdbcb549074a355"
+  },
+  "rollback": {
+    "image_id": "sha256:5a3abacbd1d5a23332e24b4f68a3459532ceff1e02418a9366ab98e8f0919c98",
+    "original_prompt_recomputed": true,
+    "expected": "red",
+    "observed": "red",
+    "passed": true,
+    "elapsed_seconds": 52.415905,
+    "qualified_fallback_restore_recheck": "pending-at-record-time"
+  },
+  "admission": {
+    "deployable": false,
+    "qualified": false,
+    "reason": "A structurally verified persistent restore failed the exact semantic oracle."
+  }
+}

sparkcache/persistent_context_cache/cache_manifest.py

@@ -2938,8 +2938,8 @@ def commit_page_snapshot(
 
         Physical extents are independent of the identity's logical chunk
         geometry. The root retains the exact 256-token boundary and chunk
-        count used by lookup/admission, while restore authenticates both each
-        extent and the reassembled byte stream before any page placement.
+        count used by lookup/admission. Restore authenticates the root metadata,
+        contiguous descriptor geometry, and each extent before placement.
         """
 
         _validate_digest(context_digest, "context_digest")

sparkcache/spark_context_cache_connector.py

@@ -3701,11 +3701,16 @@ def _load_hybrid_pages(
             )
             logger.info(
                 "spark-context-cache: SparkCache CUDA restore verified %d bytes"
-                " slabs=%d read_hash=%.1f ms submit=%.1f ms finish=%.1f ms",
+                " slabs=%d read_hash=%.1f ms placement=%.1f ms"
+                " (arena_wait=%.1f ms host_copy=%.1f ms submit_call=%.1f ms)"
+                " finish=%.1f ms",
                 result.source_bytes,
                 result.slabs,
                 result.read_and_hash_ms,
                 result.copy_and_submit_ms,
+                result.arena_wait_ms,
+                result.host_copy_ms,
+                result.submit_call_ms,
                 result.finish_ms,
             )
             return True