docs(v3): W3c oracle measured live — 1-2 ms framework overhead vs 8.4-8.7 s LLM round trip (Addendum-9)

Live rig(xai)->FlowRunner oracle-node run; self-hosted JITSON serve.rs
test-bench delta recorded; rs-graph-llm design-intent + GraphRAG-rs
references folded.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MLBnPuScZy6w9di2QEjsXM

Author: claude

.claude/board/EPIPHANIES.md

@@ -1,3 +1,8 @@
+## 2026-07-02 — E-V3-ORACLE-LIVE-1: W3c oracle node measured LIVE — graph-flow overhead is 1-2 ms against an 8.4-8.7 s LLM round trip
+**Status:** FINDING (live run, rig 0.39 xai provider, 3 calls ~$0.02; harness in session scratchpad, reproducible)
+
+Live numbers (this container, via HTTPS proxy, XAI_API_KEY): T1 direct grok-3-mini connectivity smoke 2958 ms ("OK"). T2/T3: the W3c oracle-node shape — a graph_flow::Task whose run() consults a rig agent (grok-4-0709, max_tokens 64, FailureTicket pronoun-disambiguation prompt) — executed through GraphBuilder + FlowRunner + InMemorySessionStorage: e2e 8660/8406 ms, API 8658/8405 ms, **framework overhead 2/1 ms**. Both runs Completed; the oracle answered the disambiguation correctly ("it" = the board, animacy reasoning). Verdict: the orchestration layer costs SIX orders of magnitude less than the oracle call it wraps — W3c's budget is purely the LLM round trip (seconds), reinforcing oracle-frequency-only (never per-transition) and the two-speed split. Ergonomics: the whole oracle node is ~30 lines; three frictions: rig-core's lib name is rig_core (not rig), .agent() needs the rig_core::client::CompletionClient trait in scope, errors map via graph_flow::GraphError::TaskExecutionFailed. Operator references folded: (a) rs-graph-llm IS the LangGraph-inspired Rig-integrated stateful/interruptible handler by design (author's framing; the queue-vs-state debate in that thread is answered in V3 by the kanban board being BOTH — WAL intent log + state); (b) GraphRAG-rs (automataIA/graphrag-rs) noted as RAG-side prior art with native LanceDB/Arrow support, Leiden, LightRAG, cAST tree-sitter chunking — candidate reference for the retrieval stages, not a dependency decision; (c) test-bench delta: planner serve.rs (Axum, OpenAI-compatible /v1/chat/completions, JITSON behind it) can serve as a LOCAL zero-cost oracle endpoint via rig's openai-compatible provider + graceful retry — the self-hosted langgraph test bench for CI (no external key, no spend).
+
 ## 2026-07-02 — E-V3-TEMPORAL-DEINTERLACE-1: temporal.rs is the read side of the WAL ruling — replay = a read at a pinned QueryReference
 **Status:** FINDING (operator pointer to planner temporal.rs; unifies M24 + M25 with the deinterlace machinery)
 

.claude/v3/INTEGRATION-PLAN.md

@@ -328,3 +328,20 @@ Operator: "check temporal.rs for a deeper understanding." Verified against
    dispatchable = time-admitted AND data-ready — the standing rule
    "updates reprioritize, never gate" holds because a data-blocked row is
    dropped from the PROJECTION, not refused at the writer.
+
+### Addendum-9 2026-07-02 — W3c oracle measured LIVE + the self-hosted test bench
+
+- **Live numbers** (rig xai, grok-4-0709 oracle through FlowRunner):
+  framework overhead **1-2 ms** vs **8.4-8.7 s** LLM round trip. W3c's
+  budget IS the LLM call; orchestration is free. Full data:
+  E-V3-ORACLE-LIVE-1.
+- **Test-bench delta (operator):** wire rig's OpenAI-compatible provider
+  at the planner's serve.rs (`/v1/chat/completions`, JITSON behind it)
+  as the LOCAL oracle for CI — zero external spend, no key, graceful
+  retry wrapping for a typical langchain-style bench. Lands with W3b's
+  KanbanSessionStorage tests.
+- **References:** rs-graph-llm's stated design (LangGraph-inspired,
+  Rig-integrated, interruptible-by-design, step-by-step default) matches
+  the W3b role exactly; the thread's queue-vs-stateful debate is resolved
+  in V3 by the kanban board being both (M24 WAL + state). GraphRAG-rs
+  noted as RAG prior art (native LanceDB/Arrow, Leiden, LightRAG, cAST).