Flatten graph retrieval latency

Author: gabewillen

README.md

@@ -49,6 +49,14 @@ no future or current-turn leakage, hidden labels absent, text-only RAG
 baselines. Artifact:
 `eval_runs/msc_frontier_late_200dlg_gpt55_20260630T053427Z/judge_openai_gpt55_four_system_clean.json`.
 
+A post-optimization full replay on the same 9,130-turn MSC slice preserved the
+native probe behavior exactly while flattening graph-retrieval latency. The 9
+progress checkpoints reported `GraphRetrieve.total` between 13.7 ms and 27.7
+ms, and the 9 judged probe turns reported 21.8-28.6 ms. Non-timing probe fields,
+retrieved/working memory IDs, retrieval counts, memory counts, and consolidation
+counts matched the saved baseline exactly. Artifact:
+`build/graph_profile/full_msc_verify_final/summary.json`.
+
 An earlier local blind-judge pass (2026-06-28) on a one-year sparse replay,
 judged by Gemma4-12B-AWQ over vLLM at a 131,072-token context, completed 93/93
 judgments.

docs/paper/_manuscript/index.md

@@ -3736,9 +3736,11 @@ path</td>
 <tr>
 <td>retrieval seed surface</td>
 <td>current-memory embedding lookup instead of all reconstruction
-versions</td>
-<td>latest reconstruction remains visible without unbounded scan
-growth</td>
+versions; recent seed and same-source expansion use the live retrieval
+surface cache before SQL fallback</td>
+<td>latest reconstruction remains visible without unbounded scan growth,
+and source-neighbor behavior is preserved without per-seed table
+scans</td>
 </tr>
 <tr>
 <td>graph expansion</td>
@@ -3773,6 +3775,18 @@ The important correctness point is that these are implementation changes
 to the retained engine. They do not reintroduce the removed semantic
 batch stack.
 
+On 2026-06-30, a full Meta MSC replay rerun verified the retrieval-cache
+optimization on the same 9,130-turn slice used for the hosted
+frontier-judge artifact. The run preserved native probe behavior
+exactly: all non-timing probe fields, retrieved/working memory IDs,
+retrieval counts, memory counts, and consolidation counts matched the
+saved baseline. At 1,000-event progress checkpoints,
+`GraphRetrieve.total` was 13.7-27.7 ms after optimization versus
+31.4-323.5 ms in the baseline. Across the 9 judged probe turns,
+`GraphRetrieve.total` was 21.8-28.6 ms after optimization versus
+81.6-303.6 ms in the baseline. The verification artifact is
+`build/graph_profile/full_msc_verify_final/summary.json`.
+
 ## Verification Gates
 
 The optimization and cutover work is gated by:

docs/paper/sections/11_optimization.qmd

@@ -58,7 +58,7 @@ The current branch includes the following implementation-level changes:
 |---|---|---|
 | SQLite execution | prepared-statement reuse in the store layer | same SQL semantics, lower repeated parse overhead |
 | vector ranking | bounded top-k selection and architecture SIMD kernels in the vector path | same candidate contract; SIMD accepted after output checks |
-| retrieval seed surface | current-memory embedding lookup instead of all reconstruction versions | latest reconstruction remains visible without unbounded scan growth |
+| retrieval seed surface | current-memory embedding lookup instead of all reconstruction versions; recent seed and same-source expansion use the live retrieval surface cache before SQL fallback | latest reconstruction remains visible without unbounded scan growth, and source-neighbor behavior is preserved without per-seed table scans |
 | graph expansion | retained edge types only: `co_occurs`, `similar_to`, `reinforces`, `causes`, `derived_from` | removed fact, label-bank, and temporal-retrieval side paths cannot influence rank |
 | reconstruction | knob-bounded retrieval-time reconstruction append | constructive recall remains non-decoder and can be disabled cleanly |
 | memory storage | avoid redundant object-store writes for single-signal memory payloads | durable payload identity is preserved |
@@ -67,6 +67,17 @@ The current branch includes the following implementation-level changes:
 The important correctness point is that these are implementation changes to the
 retained engine. They do not reintroduce the removed semantic batch stack.
 
+On 2026-06-30, a full Meta MSC replay rerun verified the retrieval-cache
+optimization on the same 9,130-turn slice used for the hosted frontier-judge
+artifact. The run preserved native probe behavior exactly: all non-timing probe
+fields, retrieved/working memory IDs, retrieval counts, memory counts, and
+consolidation counts matched the saved baseline. At 1,000-event progress
+checkpoints, `GraphRetrieve.total` was 13.7-27.7 ms after optimization versus
+31.4-323.5 ms in the baseline. Across the 9 judged probe turns,
+`GraphRetrieve.total` was 21.8-28.6 ms after optimization versus 81.6-303.6 ms
+in the baseline. The verification artifact is
+`build/graph_profile/full_msc_verify_final/summary.json`.
+
 ## Verification Gates
 
 The optimization and cutover work is gated by: