[P2] Live-update Learn progress card via SSE graph deltas

[AndresL230]

Problem

The Learn page's right-side Progress card pulls from graphNodes, which is fetched once at page bootstrap (frontend/src/components/screens/Learn.tsx:128,147). During a tutor session, the backend mutates mastery via services/graph_service.py::apply_graph_update, but the client never sees the new state until the page is reloaded — so the card feels stale through a session.

Proposal — cheapest "live" option

Have the tutor SSE stream (backend/routes/learn.py) emit a graph_update event whenever apply_graph_update runs, carrying just the changed node fields (id, mastery_tier, mastery_score, optionally name/course_id for new nodes). On the client, handle the event by patching the affected entries in graphNodes / graphEdges state in place.

This avoids any extra HTTP requests — the data piggybacks on the stream we are already opening for the assistant turn — and the progress card re-derives via its existing useMemo as soon as state updates.

Scope

Backend

• Extend the SSE event schema in backend/routes/learn.py with a graph_update event type.
• Capture deltas from apply_graph_update (return changed nodes/edges from the service or wrap the call) and emit them on the stream.

Frontend

• Extend the SSE consumer used by the Learn chat to recognize graph_update events.
• Apply a reducer that merges incoming node/edge deltas into graphNodes / graphEdges (upsert by id).
• No change needed to progressItems — it will recompute via useMemo.

Non-goals

• Reworking how mastery is computed.
• Refetching the full graph per turn (rejected as too heavy).
• Backfilling history for past sessions.

Acceptance

• During a live tutor session, when the model triggers a graph update, the Progress card on the right reflects the new mastery state without a page reload.
• No additional REST round-trips per turn.

[AndresL230]

Dependency note from the backend audit (epic #136): this feature relies on the tutor agent actually emitting mastery deltas, which #127 found it no longer does (only new_nodes at mastery 0.0 are emitted, never updated_nodes/mastery_delta). #127 should land first, otherwise the SSE graph_update stream proposed here would carry nothing during chat.

[AndresL230]

Coordination: the SSE transport for this is specced in #70 (stream the tutor reply token-by-token via agent.run_stream() + streamSSE). The graph_update deltas described here should ride the same stream and be emitted on the terminal done event — implement #70 and #74 together. Note #127 (agent currently emits no mastery deltas) gates this being meaningful.