docs: record canonical live multi-turn product-observation route

docs/task-context-packs.md

@@ -66,6 +66,30 @@ It makes exactly one live Ollama call per invocation unless the experiment expli
 **Rule:** During normal reasoning experiments, never create a bespoke harness, enumerate `/api/tags`,
 probe localhost, or discover/substitute another model. Use the canonical harness above.
 
+### Canonical live multi-turn product-observation route
+
+A minimal script exists at `scripts/reproduce-multi-turn-investigation.mjs`.
+
+**Execution pattern:**
+
+```bash
+# 1. Start dev server
+npx next dev
+
+# 2. Run multi-turn script (requires running dev server)
+node scripts/reproduce-multi-turn-investigation.mjs
+```
+
+**State-chaining rule:**
+
+- Start returns `situationGraph` + `selectedQuestion.question`
+- Update 1 input: `situationGraph` from start, `previousQuestion` = selected question; receives `updatedSituationGraph` as next graph
+- Update 2 input: `updatedSituationGraph` from update 1, `previousQuestion` = update 1's selected question
+- Each returned `selectedQuestion.question` becomes the next `previousQuestion`
+- If any call fails, that is enough to establish the route up to that failure — do not work around it
+
+**Maximum natural live calls:** 3 (start: 1, update 1: 1, update 2: 1)
+
 ## Pack 2 — UI and Mock Work
 
 ### Always read