graph context: a separate marker message before the first user turn

The map picks no longer glue into the opening prompt as seed text
(owner, 2026-10-04 late). POST /api/chats accepts graphContext
{propositions}; the route resolves the picks against the live graph,
derives the reasoning chains among them (chainsForSelection →
formatChainLine), and writes a graph-context custom marker as
message[0] — before the learner's first user message, in either mode.

- Core: newGraphContextMessage (model-facing content = the
  contextualization framing + proposition lines + chain lines;
  details = the structured pick); toLlmMessages flattens it to a user
  message like every marker, so the model reads it for free.
- View: projects to a user-side data-graph-context part.
- UI: GraphContextCard — minimal, muted, dashed, collapsed by default
  (counts only); expand reveals name — statement rows and the chain
  lines. Same compact-expandable pattern as the practice cards.
- Home: sends graphContext with the creation whenever picks exist;
  the prompt param carries the learner's own words only.

src/core/native-messages.ts

@@ -65,6 +65,7 @@ export interface LhMarkerMessage {
 	id: string;
 	role: "custom";
 	customType:
+		| "graph-context"
 		| "probe-handoff"
 		| "subagent-steer"
 		| "subagent-abort"
@@ -187,6 +188,42 @@ export function newPlanBlockMessage(
 export const PRACTICE_PROBLEM_TOOL = "problem";
 export const PRACTICE_CONCLUDE_TOOL = "conclude";
 
+/** One picked proposition of the graph-context marker (the learner's map
+ *  pick — name + statement as the learner SAW them at pick time). */
+export interface GraphContextPick {
+	id: string;
+	name: string;
+	statement: string;
+}
+
+/** The session's map context (owner, 2026-10-04): a SEPARATE custom
+ *  message BEFORE the learner's first user message — the picked
+ *  propositions plus the reasoning chains among them (pre-formatted
+ *  lines, formatChainLine shape). The model reads `content` (flattened
+ *  to a user message by toLlmMessages like every marker); `details`
+ *  feeds the transcript's collapsible GraphContextCard. No more seed
+ *  text glued into the first prompt. */
+export function newGraphContextMessage(picks: readonly GraphContextPick[], chains: readonly string[]): LhMarkerMessage {
+	const lines = [
+		"Contextualization — the learner picked these propositions from their knowledge map when starting this session:",
+		"",
+		"Propositions:",
+		...picks.map((p) => `- ${p.name}: ${p.statement}`),
+	];
+	if (chains.length > 0) {
+		lines.push("", "Reasoning chains among them (premises → conclusion):", ...chains.map((c) => `- ${c}`));
+	}
+	return {
+		id: newMessageId(),
+		role: "custom",
+		customType: "graph-context",
+		content: lines.join("\n"),
+		display: true,
+		details: { propositions: picks, chains },
+		timestamp: Date.now(),
+	};
+}
+
 /** The learner's thinking submission (the practice dock's two buttons):
  *  a user-register message the model reads WITH its label (hint vs
  *  commit — the register is the whole difference), while `details` keeps

src/core/view.ts

@@ -290,6 +290,17 @@ function applyResultTo(part: LhViewPart | undefined, result: ToolResultMessage):
 /** A marker's projected message (handoff/steers) or splice (plan block). */
 function projectMarker(marker: LhMarkerMessage, out: LhViewMessage[]): void {
 	switch (marker.customType) {
+		case "graph-context": {
+			// The session's map context (owner, 2026-10-04): user-side —
+			// a minimal collapsible card ABOVE the first user message;
+			// never a text bubble (scaffolding, like the thinking card).
+			out.push({
+				id: marker.id,
+				role: "user",
+				parts: [{ type: "data-graph-context", id: `gc_${marker.id}`, data: marker.details }],
+			});
+			return;
+		}
 		case "probe-handoff": {
 			out.push({
 				id: marker.id,